title stringlengths 1 300 | score int64 0 8.54k | selftext stringlengths 0 41.5k | created timestamp[ns]date 2023-04-01 04:30:41 2026-03-04 02:14:14 ⌀ | url stringlengths 0 878 | author stringlengths 3 20 | domain stringlengths 0 82 | edited timestamp[ns]date 1970-01-01 00:00:00 2026-02-19 14:51:53 | gilded int64 0 2 | gildings stringclasses 7
values | id stringlengths 7 7 | locked bool 2
classes | media stringlengths 646 1.8k ⌀ | name stringlengths 10 10 | permalink stringlengths 33 82 | spoiler bool 2
classes | stickied bool 2
classes | thumbnail stringlengths 4 213 ⌀ | ups int64 0 8.54k | preview stringlengths 301 5.01k ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
PLX/PEX PCIe 4.0 seems to help for LLMs and P2P! I.e. PEX88096 (1 PCIe 4.0 X16 to 5 PCIE 4.0 X16) and others, and comparison vs bifurcation. | 39 | Hello guys, hoping you're having a good day.
I do this post if it helps for information for some users that don't know about switches.
Before anything, I have all the switches I mention on this post but the PCIe 5.0 ones and PEX88080 one. All bought from aliexpress, and all working fine, ranging from 100 to 500USD. If you're interested in the links let me know!
Also, English isn't my first language, so if you found something not written correctly also let me know!
What are PCIe switches?
PCIe switches like the Broadcom PEX88000 (Gen4) and PEX89000 (Gen5) series are essentially **packet-routing fabrics** for PCIe. They're non-transparent bridges that create a hierarchical PCIe topology, allowing multiple downstream devices to share one or more upstream ports connecting to the CPU's root complex.
Think of them as Ethernet switches but for PCIe packets. They contain:
* One or more **upstream ports** (connecting toward the CPU)
* Multiple **downstream ports** (connecting to endpoints like GPUs)
* An internal **crossbar switch fabric** that routes TLP (Transaction Layer Packets) between ports
For example one of them looks the one of the picture, also some ones look like this:
[X16 4.0 upstream via dual SlimSAS 8i uplink to 4\*X16 4.0 slots + 2 SlimSAS 8i downstream](https://preview.redd.it/c0gd2u9jdr8g1.png?width=960&format=png&auto=webp&s=9e4f49e6e908c5a04761a53df1b617af73c6c584)
What are some other benefits of switches?
* You don't need PCIe bifurcation motherboard support, the PLX/PEX switch inside does everything.
* So for example you can split a X4 slot into X1/X1/X1/X1, or X2/X1/X1, etc and dynamically, those limits will happen when you use everything fully at the same time.
* It works out of the box, you can boot on drives attached to them, and for either OS Linux or Windows.
* As PCIe is birectional, it helps a lot for P2P.
Would you wonder, how do they create so many slots from a single one?
You don't magically get more bandwidth than the slot offers (i.e. 32 GiB/s bidirectional), but if you use 2 PCIe 4.0 slots on that switch for example, you could get about 64GiB/s total if you write to one side and read from the other.
The switch presents multiple independent downstream ports (say, 4× x16 slots), each appearing as a separate PCIe link to the devices attached.
When GPU-A sends a TLP to system memory, the switch routes it through the crossbar to the upstream port. When GPU-B does the same, traffic is interleaved/arbitrated. The switch handles flow control, credit management, and QoS.
So then, traffic between downstream ports (GPU-to-GPU P2P) can traverse the switch fabric **without going through the upstream port at all**. This is why switches are valuable for multi-GPU—you could get full local bandwidth for P2P transfers.
Another switch example are these ones:
**PEX88024 (PCIe 4.0 X8 to 4 PCIe 4.0 X4 M2)**
[PEX88024 Switch](https://preview.redd.it/rkwzvadher8g1.png?width=800&format=png&auto=webp&s=5e9c6749d058f7807821f36ac4b8b1764dcee6da)
**PLX88048 (PCIe 4.0 X16 to 8 PCIe 4.0 X4 M2 and 2 SlimSAS 8i to 2x 4i each)**
[PLX88048 Switch](https://preview.redd.it/r2e6r46ser8g1.png?width=2632&format=png&auto=webp&s=08550c17b3cc9030f23eba9136b7aa82cbfb0973)
**PEX88048 variant: PCIE 4.0 X16 to 4 SlimSAS 8i (or 4x8 PCIe 4.0). In this one you can do either X16/X16, X8/X8/X8/X8, or X4/X4/X4/X4/X4/X4/X4/X4.**
[PEX88048 Switch](https://preview.redd.it/thoa2y33fr8g1.png?width=800&format=png&auto=webp&s=b985311a163b62c984aad4ee08c9f37e19f182fb)
PEX88080 (X16 4.0 to 4\*X16 4.0 slots)
[PEX88080 Switch](https://preview.redd.it/gadycyqffr8g1.png?width=1000&format=png&auto=webp&s=7d4e4de1add14208c0399c53b40115dc3468b392)
PLX88096 (Already shown one on the start, so here it is another one). PCIe X16 4.0 to 10 SlimSAS 8i ports: Supports 5\*X16 4.0, or 10\*X8 4.0, or 20\*X4 4.0.
[PEX88096 Switch](https://preview.redd.it/1yhj5ud4gr8g1.png?width=800&format=png&auto=webp&s=7d43c727d23143aa8487e007fca7554585e5eb67)
PEX89048: PCIe 5.0 X16 uplink to 4xMCIO 8i ports (so you can do X16/X16 5.0, or X8/X8/X8/X8 5.0, or 8\*X4 5.0)
[Rocket 1628A, PEX89048 Switch](https://preview.redd.it/mf1dp5qegr8g1.png?width=1065&format=png&auto=webp&s=16d9d8d4b277d3d1824a23c5a82cd26da9789243)
So what are the demerits for something that sounds so good?
* It is expensive, like a LOT more expensive than bifurcation cards.
* It add latency in the ns terms, which may or not affect your workload.
* Requires something external on your PC vs just enabling bifurcation on your motherboard BIOS.
A good table comparison would be:
# PCIe Switch vs. Bifurcation
|Aspect|Bifurcation|PCIe Switch|
|:-|:-|:-|
|**What it is**|CPU/chipset configuration that splits a single physical slot's lanes|Active silicon device with its own logic|
|**Hardware**|No additional hardware (just BIOS setting)|Requires switch chip ($$$)|
|**Bandwidth**|Divides lanes statically (x16 → 2×8, 4×4, etc.)|Shares bandwidth dynamically via arbitration|
|**Device visibility**|Each bifurcated segment is a direct CPU link|Devices sit behind switch in topology hierarchy|
|**P2P traffic**|Must traverse CPU root complex|Can route locally within switch fabric|
|**Latency**|Lower (direct to root complex)|Slightly higher (extra hop through switch)|
|**Flexibility**|Fixed by BIOS/physical slot|Can be reconfigured, supports hot-plug|
|**Cost**|Free|Significant (switch chips are expensive)|
# Practical Example
**Bifurcation scenario**: Your motherboard has an x16 slot. You set BIOS to 4×4 bifurcation and use a passive riser to install four NVMe drives. Each drive gets a dedicated x4 link straight to the CPU, but you've "spent" 16 lanes from your CPU's lane budget.
**Switch scenario**: You have an x16 slot connected to a PEX88096 card. That card provides 4× x16 downstream slots (64 lanes downstream from 16 upstream). Four GPUs can each negotiate x16 links. They share the x16 upstream bandwidth to CPU, but GPU-to-GPU P2P gets full switch fabric bandwidth (no CPU bottleneck). You've still only "spent" 16 CPU lanes.
\--------------------------
Hope you found this post informative! Any question is welcome. | 2025-12-22T14:01:51 | https://www.reddit.com/r/LocalLLaMA/comments/1pt0av6/plxpex_pcie_40_seems_to_help_for_llms_and_p2p_ie/ | panchovix | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pt0av6 | false | null | t3_1pt0av6 | /r/LocalLLaMA/comments/1pt0av6/plxpex_pcie_40_seems_to_help_for_llms_and_p2p_ie/ | false | false | 39 | null | |
Why does this feel more stable than a normal LLM chat? | 1 | [removed] | 2025-12-22T13:57:32 | QualityEvery6965 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pt075k | false | null | t3_1pt075k | /r/LocalLLaMA/comments/1pt075k/why_does_this_feel_more_stable_than_a_normal_llm/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': 'jh1tc3tbhr8g1', 'resolutions': [{'height': 134, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=108&crop=smart&auto=webp&s=a7b1211520ca4b39d8169b810299893f354204e1', 'width': 108}, {'height': 268, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=216&crop=smart&auto=webp&s=882067f6da7a5b3712bd3319dc7d4c767cf7a1d1', 'width': 216}, {'height': 397, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=320&crop=smart&auto=webp&s=44303862fcc848aad07627d1ef66692824b3e219', 'width': 320}, {'height': 794, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=640&crop=smart&auto=webp&s=a554c3570a627fe55f66175efad0f290a4e685ed', 'width': 640}, {'height': 1192, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=960&crop=smart&auto=webp&s=d7d006c7699faa62049baec3007e28122e7d6067', 'width': 960}, {'height': 1341, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?width=1080&crop=smart&auto=webp&s=4f42fa12f10d6d3e7162e6379c127cdd8e6a7650', 'width': 1080}], 'source': {'height': 3756, 'url': 'https://preview.redd.it/jh1tc3tbhr8g1.jpeg?auto=webp&s=48bac6ac1bc8acb0b632233b042636f95b940784', 'width': 3024}, 'variants': {}}]} | |
Agent builders/devs. What are the most frustrating unsolved problems you experience when building agents ? | 0 | Let me hear your thoughts guys I’m all ears 👂
I’ve been talking a lot lately to agent developers because I’ve been building with some langchain users a tool called Scope (link in the comments ).
And we’re struggling to understand if the issue we wish to resolve, which in one sentence is :
“help agent developers quickly see where and why an LLM agent run went wrong, by turning raw traces into a clear decision story with concrete fix suggestions.”
I hope that way of saying it is understandable.
So as said previously we’re struggling to understand if we are solving the problem correctly, obviously right now scope is just a demo version we will implement SDK integrations to get more data on agent runs and try to deduce more insights because with just JSON traces were kinda limited, but the real question is : is that problem already solved if not what’s the real problem to solve and if you have an answer for that one then how do we solve that said problem?
Let me know 💪 | 2025-12-22T13:57:07 | https://www.reddit.com/r/LocalLLaMA/comments/1pt06tv/agent_buildersdevs_what_are_the_most_frustrating/ | AdVivid5763 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pt06tv | false | null | t3_1pt06tv | /r/LocalLLaMA/comments/1pt06tv/agent_buildersdevs_what_are_the_most_frustrating/ | false | false | self | 0 | null |
[Tool] imesde: Zero-GPU, In-Memory Vector Engine for Real-Time Local RAG | 6 | Hi r/LocalLLaMA,
I wanted to share a project I’ve been working on called `imesde` (In-MEmory Streaming Data Engine).
If you’ve ever tried to build a RAG system for live data streams (logs, tweets, stock tickers, or even a fast Twitch chat), you’ve probably hit the "Knowledge Gap": traditional Vector DBs are great for persistence but they are often too slow to index in real-time or too disk-heavy for what is essentially "short-term memory."
`imesde` is designed to be the "Semantic Pipe" for your local LLMs. It lives entirely in RAM, runs on CPU, and has zero indexing latency.
# 🧠 Why should you care?
* **Zero-GPU Dependency**: `imesde` is intentionally architected for CPUs. It uses SIMD-accelerated dot product kernels to process Int8 Quantized models.
* **The "Infinite Window" RAG**: Instead of a static database, it’s a Sharded Circular Buffer. As new data flows in, old data flows out. Your LLM always stays focused on the most relevant.
* **Zero Indexing Latency**: The moment a packet arrives, it's available for retrieval.
* **Local-First & Private**: It uses in-process vectorization via ONNX. Your data never leaves your machine, making it perfect for private SRE monitoring or local agents.
# 🚀 Performance Benchmark on a 5000 records dataset (Apple M4)
* **Avg Search Latency**: 232.74 μs
* **P99 Search Latency**: 383.75 μs
* **Engine Throughput**: \~4,300 queries/sec
* **Avg Embedding Time**: 1.93 ms (using bge-small-en-v1.5 int8)
# 🐍 Minimal Python Snippet
import imesde
# Initialize with a local ONNX model
db = imesde.PyImesde(
"model/model.onnx",
"model/tokenizer.json",
num_shards=32,
shard_size=2048
)
# Ingest live stream (massively parallelized via Rust)
db.ingest_batch(["User login at 10:00", "DB Query timeout", "High latency detected"])
# Retrieve the absolute latest context for your LLM
context = db.search("database issues", k=3)
# -> results: [(text, score), ...]
# Showcase: The Semantic Radar
I've included an example in the repo that monitors the global aviation firehose (OpenSky API).
It ingests \~10,000 aircraft states per minute, maps them to semantic concepts (e.g., "dangerous high speed at low altitude"), and triggers a local LLM (via Ollama) only when a real anomaly is detected.
I’m looking for community feedback on two fronts: Which integrations would make this most useful in your local workflows? Also, do you have any favorite ultra-lightweight ONNX models to recommend for this type of high-speed CPU streaming?
GitHub: [https://github.com/imesde/imesde](https://github.com/imesde/imesde)
Built with Rust. MIT Licensed. | 2025-12-22T13:44:33 | https://www.reddit.com/r/LocalLLaMA/comments/1pszwoq/tool_imesde_zerogpu_inmemory_vector_engine_for/ | alessiopelliccione | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pszwoq | false | null | t3_1pszwoq | /r/LocalLLaMA/comments/1pszwoq/tool_imesde_zerogpu_inmemory_vector_engine_for/ | false | false | self | 6 | {'enabled': False, 'images': [{'id': 'uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=108&crop=smart&auto=webp&s=24cf33d6e0540ee7d27b0f6eeaab365a76fd1779', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=216&crop=smart&auto=webp&s=be13fe53814dcd3136aac18e785c28d63a279f44', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=320&crop=smart&auto=webp&s=4c8f7fa95377ee5738f96e0fc113fde4415a6b4f', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=640&crop=smart&auto=webp&s=7ab6fdb526369770e5117216160657d2be824874', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=960&crop=smart&auto=webp&s=ddc340d333df327755e1acf2c66d83b278260b0a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?width=1080&crop=smart&auto=webp&s=3fbe7c2a74da928ae48be0207191db340047001c', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/uKpKFUAZri4LO9ESYSMUGQVgnfXaA_xzIVPLFs4vg5g.png?auto=webp&s=50a3016c64a3f622632957cf81754488b17a5584', 'width': 1200}, 'variants': {}}]} |
using gemma for perception and normalization in agents? | 2 | If we use gemma 314B or 27B as a pre processor to responses received from tools, clean it as needed and feed only necessary info into the core llm for long agentic workflows and jobs, would we have much high accuracy and reliability? Are people using gemma in agents? | 2025-12-22T13:40:39 | https://www.reddit.com/r/LocalLLaMA/comments/1psztmt/using_gemma_for_perception_and_normalization_in/ | LankyLychee8916 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psztmt | false | null | t3_1psztmt | /r/LocalLLaMA/comments/1psztmt/using_gemma_for_perception_and_normalization_in/ | false | false | self | 2 | null |
72Gb VRAM (3x 3090) / 128Gb DDR4 / Mylan CPU What code model can I test? | 0 | Hey guys, just landed into the world of LocalLLM,
My rigs is small compare to some (planning on going up to 8x 3090 Q1 2026) however I'd like to test the capability on the 3x 3090, what model / setup have you run that for code specific?
I want to test this with Roocode and the bomerang feature - the goal would be to use as much compute as possible for an extensive session and see what the rig can deliver on his own
suggestion are welcome :)
Quentin | 2025-12-22T13:39:19 | https://www.reddit.com/r/LocalLLaMA/comments/1pszsn1/72gb_vram_3x_3090_128gb_ddr4_mylan_cpu_what_code/ | shvz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pszsn1 | false | null | t3_1pszsn1 | /r/LocalLLaMA/comments/1pszsn1/72gb_vram_3x_3090_128gb_ddr4_mylan_cpu_what_code/ | false | false | self | 0 | null |
Tracking LLM Benchmarks Across arXiv | 1 | alphaXiv indexed results from millions of papers to surface which models are SOTA according to open source research.
[https://www.alphaxiv.org/state-of-the-art](https://www.alphaxiv.org/state-of-the-art) | 2025-12-22T13:39:03 | https://v.redd.it/5mug8gaedr8g1 | Dear-Success-1441 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pszsgd | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/5mug8gaedr8g1/DASHPlaylist.mpd?a=1769002760%2CYTdmOTVmZDRiMjdlODBlMmFkZDBhMzhiOGIzZThjZjZiY2JhNjM2MzQ5YmFiZWI2Y2MwZDMxYzMzNGE3ZWM0Yw%3D%3D&v=1&f=sd', 'duration': 17, 'fallback_url': 'https://v.redd.it/5mug8gaedr8g1/CMAF_720.mp4?source=fallback', 'has_audio': False, 'height': 708, 'hls_url': 'https://v.redd.it/5mug8gaedr8g1/HLSPlaylist.m3u8?a=1769002760%2CYzE5YTdiYzM0YmU2M2Y4ZjA2NTIwODYzYjk4OTEyMTRkZDIxMTRjNTM2MWQzYjZhZWRhOGMxN2VlZTJjZTZjMQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/5mug8gaedr8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1280}} | t3_1pszsgd | /r/LocalLLaMA/comments/1pszsgd/tracking_llm_benchmarks_across_arxiv/ | false | false | 1 | {'enabled': False, 'images': [{'id': 'MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID', 'resolutions': [{'height': 59, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=108&crop=smart&format=pjpg&auto=webp&s=f18b67735df3b516b4939055db81c8aecc0c29c0', 'width': 108}, {'height': 119, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=216&crop=smart&format=pjpg&auto=webp&s=712176a9a9758ad527236f05b23450649c47c0e6', 'width': 216}, {'height': 177, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=320&crop=smart&format=pjpg&auto=webp&s=b90515cadaaaf7620dc5e75ca0ad15161dbdad66', 'width': 320}, {'height': 354, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=640&crop=smart&format=pjpg&auto=webp&s=87ec9e67e6de5c61395acf5bdee2fdac5f426f24', 'width': 640}, {'height': 531, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=960&crop=smart&format=pjpg&auto=webp&s=cdc643543b48ab2dbcb9f04f7d1137605c98b6d9', 'width': 960}, {'height': 598, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?width=1080&crop=smart&format=pjpg&auto=webp&s=67c9e719690e02945575b8a7968c722ec871774d', 'width': 1080}], 'source': {'height': 720, 'url': 'https://external-preview.redd.it/MDBsajBvYWVkcjhnMS29ePEkxjn8ArZXOta0FrRD7gYg5O_CjmkWTyFY5DID.png?format=pjpg&auto=webp&s=5ddcfb2ada62c157b42b69c8afe59e286e1ae6c1', 'width': 1300}, 'variants': {}}]} | |
Upgrade in 2026 or wait until the price of a new >1000GB/s 768GB VRAM/URAM machine is <=$4k? | 0 | Judging the state of the VRAM and Mac market, we would need to wait another 5-9 years to get a brand new >1000GB/s 768GB VRAM/URAM machine for less than 4k. Maybe it will be even worse, in 5-9years, so many people will have lost their jobs and the enterprise market will have become so profitable and big, that GPU and Ram providers will stop selling to consumers and consumer PC providers. Maybe some of us should just buy a macbook/ Mac Studio/ ai max 395 in 2026 and run a <=108b model locally, and we should run >200b models in the cloud instead.. Or should we be content running small models and using the API/sub?
IT is looking like GPUs are gonna get more expensive since nvidia is reducing the supply... Unless the ai bubble pops, it will remain expensive. Even if it pops, maybe the government will pump more money into it to keep it afloat.. And DS v4 or kimi k3 will likely be even bigger than v3.2 and k2.
IT seems like the dream of running the newest best q6-8 deepseek v.# or q4/8 kimi k# thinking at >=28tk/s locally will never happen.. Even running minmax 2.1 q6 locally at >=40TK/s will be unattainbale unless you spend more than 5.5k.. | 2025-12-22T13:28:00 | https://www.reddit.com/r/LocalLLaMA/comments/1pszk1l/upgrade_in_2026_or_wait_until_the_price_of_a_new/ | power97992 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pszk1l | false | null | t3_1pszk1l | /r/LocalLLaMA/comments/1pszk1l/upgrade_in_2026_or_wait_until_the_price_of_a_new/ | false | false | self | 0 | null |
Looking for a stop motion / frame sequencing workflow for Wan 2.2? | 2 | Will you require a specific lora or a just settings variance/nodes to produce a stop motion 8 frame per second videos? Or a frame sequencing set up?
If not, can you make them via Z-image, Qwen edits...?
Thanks. | 2025-12-22T13:26:04 | https://www.reddit.com/r/LocalLLaMA/comments/1pszimg/looking_for_a_stop_motion_frame_sequencing/ | Slight_Tone_2188 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pszimg | false | null | t3_1pszimg | /r/LocalLLaMA/comments/1pszimg/looking_for_a_stop_motion_frame_sequencing/ | false | false | self | 2 | null |
Building a Local LLM for Homeschooling on an Old i5-3570—Any Advice? | 2 | I’m a homeschooling parent, and—purely out of curiosity and privacy concerns—I’m trying to run a local LLM on my aging desktop: an Intel i5-3570 with 16GB RAM. Yeah, I know it sounds ridiculous in 2025, but I really don’t want to send my kids’ learning profiles or personal data to any public/cloud API.
My goal is to build a helper that can:
- Draft lesson plans & curriculum outlines
- Act as a discussion partner (Socratic-style, ideally)
- Generate worksheets and exams
- Help assess both academic progress and behavior
I’m planning to use RAG to ground it in our actual curriculum materials (PDFs, notes, etc.), so accuracy matters more than speed—which is good, because speed definitely won’t be happening on this rig.
Current stack: llama.cpp + Open WebUI, running Qwen3-4B-Instruct-2507. It works… but feels too rigid, and worse—it often can’t find answers that definitely exist in my reference files, especially when the info is inside PDFs with tables. The RAG retrieval just misses it, even though the document was uploaded and appears indexed.
I’d love alternative suggestions:
- Other small models that handle RAG better on CPU-only setups?
- PDF preprocessing tricks to make tables/text more RAG-friendly?
- Or just general “you’re not crazy for trying this” validation? 😅
P.S. Yes, I used AI to write this post—my English sucks, lol. | 2025-12-22T13:17:17 | https://www.reddit.com/r/LocalLLaMA/comments/1pszbz6/building_a_local_llm_for_homeschooling_on_an_old/ | ManufacturerLive6214 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pszbz6 | false | null | t3_1pszbz6 | /r/LocalLLaMA/comments/1pszbz6/building_a_local_llm_for_homeschooling_on_an_old/ | false | false | self | 2 | null |
Devstral 2 vs Devstral Small 2: A 30-Minute Playground Test for Multi-File Coding Tasks | 1 | [removed] | 2025-12-22T13:06:28 | https://www.reddit.com/r/LocalLLaMA/comments/1psz44k/devstral_2_vs_devstral_small_2_a_30minute/ | Unique_Can_2569 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psz44k | false | null | t3_1psz44k | /r/LocalLLaMA/comments/1psz44k/devstral_2_vs_devstral_small_2_a_30minute/ | false | false | self | 1 | null |
Kimi K2 Thinking is the least sycophantic open-source AI, according to research by Anthropic | 75 | 2025-12-22T12:58:00 | https://www.reddit.com/r/LocalLLaMA/comments/1psyxwo/kimi_k2_thinking_is_the_least_sycophantic/ | InternationalAsk1490 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psyxwo | false | null | t3_1psyxwo | /r/LocalLLaMA/comments/1psyxwo/kimi_k2_thinking_is_the_least_sycophantic/ | false | false | 75 | null | ||
upstage/Solar-Open-100B · Hugging Face | 115 | ...do you remember [https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0) from 2024?
It looks like they have something new:
# [](https://huggingface.co/upstage/Solar-Open-100B#solar-open)Solar Open
**Solar Open** is Upstage's flagship **102B-parameter** large language model, trained **entirely from scratch** and released under the **Solar-Apache License 2.0** (see [LICENSE](https://huggingface.co/upstage/Solar-Open-100B/blob/main/LICENSE)). As a **Mixture-of-Experts (MoE)** architecture, it delivers enterprise-grade performance in reasoning, instruction-following, and agentic capabilities—all while prioritizing transparency and customization for the open-source community.
# [](https://huggingface.co/upstage/Solar-Open-100B#highlights)
# Highlights
* **MoE Architecture (102B / 12B):** Built on a Mixture-of-Experts architecture with **102B total / 12B active parameters**. This design delivers the knowledge depth of a massive model with the inference speed and cost-efficiency of a much smaller model.
* **Massive Training Scale:** Pre-trained on **19.7 trillion tokens**, ensuring broad knowledge coverage and robust reasoning capabilities across various domains.
# [](https://huggingface.co/upstage/Solar-Open-100B#model-overview)
# Model Overview
* **Model Name:** Solar Open 100B
* **Hugging Face ID:** Upstage/Solar-Open-100B
* **Architecture:** Mixture-of-Experts (MoE)
* **Total Parameters:** 102.6B
* **Active Parameters:** 12B (per token)
* **Experts:** 129 Experts (top 8 among 128 Routed + 1 Shared)
* **Pre-training Tokens:** 19.7 Trillion
* **Context Length:** 128k
* **Training Hardware:** NVIDIA B200 GPUs
* **License:** **Solar-Apache License 2.0** (See [LICENSE](https://huggingface.co/upstage/Solar-Open-100B/blob/main/LICENSE)) | 2025-12-22T12:47:16 | https://huggingface.co/upstage/Solar-Open-100B | jacek2023 | huggingface.co | 1970-01-01T00:00:00 | 0 | {} | 1psyqha | false | null | t3_1psyqha | /r/LocalLLaMA/comments/1psyqha/upstagesolaropen100b_hugging_face/ | false | false | default | 115 | {'enabled': False, 'images': [{'id': 'KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=108&crop=smart&auto=webp&s=9bb4811c12a9d4a0051defb855ee79a97b25f439', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=216&crop=smart&auto=webp&s=e40e112b4377d8cd33fb4e132623edf103a24690', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=320&crop=smart&auto=webp&s=78ce7ec06f10da616459aa5a38a38f52eb028d5a', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=640&crop=smart&auto=webp&s=559420850549bc9f134f2c1908f869fba5d48dbf', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=960&crop=smart&auto=webp&s=eceab33ce7979c5687a48ff76aed0b56e7a438e6', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?width=1080&crop=smart&auto=webp&s=4de5b623e2ec1833e25981cfeea197346abc9f95', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/KGZNZzWd5K6vM05pplkzhPrPxHhbMAP1w-s6MnLTkhM.png?auto=webp&s=356d2a4fed734022049b5fc170ffa741ba162b91', 'width': 1200}, 'variants': {}}]} |
Built a small client-side LLM preprocessor to reduce API token costs — looking for honest feedback | 1 | [removed] | 2025-12-22T12:45:34 | https://www.reddit.com/r/LocalLLaMA/comments/1psypb7/built_a_small_clientside_llm_preprocessor_to/ | Admirable-Degree2876 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psypb7 | false | null | t3_1psypb7 | /r/LocalLLaMA/comments/1psypb7/built_a_small_clientside_llm_preprocessor_to/ | false | false | self | 1 | null |
Best model choice | 2 | Hi all. I’m looking for advice on what model is best for my project:
Custom RAG protocol to summarise regulatory requirement held across various government laws/regs/guidance.
I’m only working with 8GB of VRAM at the moment and 32GB of RAM but ideally I’d like to avoid offloading to manage latency.
Context window is key as I will likely be feeding it 5-10 chunks of up to 400 tokens each and it needs to handle that. The questions will be simple and direct without much need for semantic understanding.
It’s a proof of concept at the moment so if it works I’ll upgrade my rig. Equally it doesn’t have to be perfect as I’m aware of the limitations. | 2025-12-22T12:45:19 | https://www.reddit.com/r/LocalLLaMA/comments/1psyp5n/best_model_choice/ | Dank-but-true | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psyp5n | false | null | t3_1psyp5n | /r/LocalLLaMA/comments/1psyp5n/best_model_choice/ | false | false | self | 2 | null |
Same LLMs, fewer prompts — why does this feel more reliable? | 0 | Most AI chats today rely on:
• prompt-based personality
• implicit assumptions
• probabilistic drift over time
I tried the opposite approach:
• architecture-defined behavior
• explicit state and routing
• treating the LLM as a renderer, not a decision-maker
Unexpectedly, the system became more logical and predictable, not less.
Screenshot attached shows a trivial example (language switch mid-conversation), but the pattern holds across longer sessions.
I’m not selling anything — genuinely curious what people here think:
Is this just an illusion caused by tighter constraints, or is “chat-first AI” fundamentally unstable by design? | 2025-12-22T12:35:01 | QualityEvery6965 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psyi5o | false | null | t3_1psyi5o | /r/LocalLLaMA/comments/1psyi5o/same_llms_fewer_prompts_why_does_this_feel_more/ | false | false | default | 0 | {'enabled': True, 'images': [{'id': '9rejj1se2r8g1', 'resolutions': [{'height': 134, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=108&crop=smart&auto=webp&s=50f9a047547925e99c9a90236dc9ca1d9f5d03b3', 'width': 108}, {'height': 268, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=216&crop=smart&auto=webp&s=f6c8484489cacca9b34ff786e2e4cbb0a3999922', 'width': 216}, {'height': 397, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=320&crop=smart&auto=webp&s=7b3d22605d49ed680dcf1bfa060bd67e9997e1a9', 'width': 320}, {'height': 794, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=640&crop=smart&auto=webp&s=30e56f78f8ea6703f1977e34e16ed7467581685a', 'width': 640}, {'height': 1192, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=960&crop=smart&auto=webp&s=8e8c7e2d0cdd2adb3dbbea372c69a3fe5a03be6c', 'width': 960}, {'height': 1341, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?width=1080&crop=smart&auto=webp&s=2005c0cb1615dfa4d9ca9edecf87483122861d99', 'width': 1080}], 'source': {'height': 3756, 'url': 'https://preview.redd.it/9rejj1se2r8g1.jpeg?auto=webp&s=137a485160d61dcb154ef0092753fcccfa25d623', 'width': 3024}, 'variants': {}}]} | |
[User research] Building a long-running single/multi-agent system — what real workflows should it handle? (free credits) | 0 | Hey all — disclosure: I’m one of the builders of an **agent system** that supports **long-running tasks** (multi-day / stateful) and **single or multi-agent workflows**.
I’m running a **use-case collection** to pressure-test what people actually want and where existing agents fail.
**What I’m looking for:**
* A workflow you wish an agent could do end-to-end (multi-step, multi-tool, possibly over days)
* Or a workflow you tried with agents and it failed (hallucination, wrong tool usage, missing verification, etc.)
**What you get:**
* I’ll provide **trial credits** for selected use-cases (I’ll run this in batches; credits are limited per batch).
**Please comment using this template (so it’s implementable):**
1. Goal (what “done” looks like)
2. Inputs (files/URLs/APIs? any privacy constraints)
3. Constraints (time/budget/environment/rules)
4. What must be verified (evidence/logs/checks you’d want)
5. Expected output format (report/CSV/code/tickets/dashboard)
6. Common failure modes you’ve seen (if any)
If you can’t share details publicly, post an abstract version (Goal + Constraints + Output) and I’ll ask minimal follow-ups in comments.
| 2025-12-22T12:31:50 | https://www.reddit.com/r/LocalLLaMA/comments/1psyg1s/user_research_building_a_longrunning/ | PerceptionAble5423 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psyg1s | false | null | t3_1psyg1s | /r/LocalLLaMA/comments/1psyg1s/user_research_building_a_longrunning/ | false | false | self | 0 | null |
Built a safe and simple way to connect agents to your data stack | 0 | Here’s how Pylar works in simple terms.
First, you connect Pylar to where your data already lives, using pre-built integrations to databases, warehouses, CRMs, product and support tools. No rebuilding, no moving data around.
Next, you create **agent data views**. Think of these as clean, task-specific snapshots of your data. You decide what fields are included, what’s hidden, and how things are joined. This is the *only* data an agent is allowed to see.
On top of these views, you can setup a custom **MCP layer**. This is how agents actually access data. Instead of running raw queries, agents call tools that sit on top of these views. The MCP layer enforces rules, guardrails and policies - what can be queried, how often, and within what limits.
Before anything goes live, you test everything in the **agent playground**. You can see exactly how the agent uses each tool, what data it touches, and what responses it generates. This helps catch issues early, before real users or workflows are involved.
You can deploy this entire package to any agent builder you're using today - n8n, agno, LangGraph, Zapier etc- via a single link.
Once deployed, Pylar runs **evals** on agent behavior - tracking usage, failures, retries, and query patterns. This makes it easy to understand whether an agent is behaving as expected or drifting over time.
In short:
Pylar helps you connect your data, shape it for agents, control access through MCP tools, test safely in a playground, and monitor behavior after deployment — all so agents stay useful, safe, and predictable.
Would love to have feedback on what we've built. | 2025-12-22T12:12:46 | https://v.redd.it/shus72ijyq8g1 | Better-Department662 | /r/LocalLLaMA/comments/1psy3at/built_a_safe_and_simple_way_to_connect_agents_to/ | 1970-01-01T00:00:00 | 0 | {} | 1psy3at | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/shus72ijyq8g1/DASHPlaylist.mpd?a=1769127174%2CODU2YTRjMjQ2ZDdhZTlkMmNmZWJhM2JjNjI0Nzc2Y2Q4YWMzYzZjOTdlMGZjY2E5MTUzYjcxYjA3MWI2ZDIwMA%3D%3D&v=1&f=sd', 'duration': 105, 'fallback_url': 'https://v.redd.it/shus72ijyq8g1/CMAF_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/shus72ijyq8g1/HLSPlaylist.m3u8?a=1769127174%2CZGM4ZmQ5ODhhMDFhZWU2NTc5NmI2NWFiNmVhNmU5ZjMyNjgwMmE3NWIwMWQzNGQ1NTQ1MjVjYmFjNTdiOGFjZg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/shus72ijyq8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1728}} | t3_1psy3at | /r/LocalLLaMA/comments/1psy3at/built_a_safe_and_simple_way_to_connect_agents_to/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU', 'resolutions': [{'height': 67, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=108&crop=smart&format=pjpg&auto=webp&s=e390b0ba7417e1560174758acb2c18b738cc78e8', 'width': 108}, {'height': 135, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=216&crop=smart&format=pjpg&auto=webp&s=faed140c4546c67277c1e19a460841473d10d55f', 'width': 216}, {'height': 200, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=320&crop=smart&format=pjpg&auto=webp&s=37ffa54efefe7b83c16d079b6b4a209b6b2f39db', 'width': 320}, {'height': 400, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=640&crop=smart&format=pjpg&auto=webp&s=9e730dd4c5fc5e1e66afb0434238bf29647bd336', 'width': 640}, {'height': 600, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=960&crop=smart&format=pjpg&auto=webp&s=0c0ff6df94120c2eafdbd2d4fed2843bc6507e17', 'width': 960}, {'height': 675, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?width=1080&crop=smart&format=pjpg&auto=webp&s=0aeacdf9ad199fbadca05d9489099d69e2d49216', 'width': 1080}], 'source': {'height': 2160, 'url': 'https://external-preview.redd.it/enFjMXJhbGp5cThnMWRa4OkQHDhkf_86kr5Q8OcB2_phR992cHwluPjyGLdU.png?format=pjpg&auto=webp&s=30fb062a8d677d9d3a75104027ce94cfba17b168', 'width': 3456}, 'variants': {}}]} | |
Qwen3-235B-W4A16 is S tier | 19 | The best model I’ve tried that can fit on my 200GB VRAM is surprisingly this Qwen model.
It is able to navigate complex puzzles, and agentic environments with grace, it’s a good coder, and extremely fast.
I am even comparing this to remotely hosted FP16 models like GLM-4.6
This might be old news but I wrote of Qwen since I typically dislike how locked down their models feel. Hope this is helpful. | 2025-12-22T12:12:42 | https://www.reddit.com/r/LocalLLaMA/comments/1psy387/qwen3235bw4a16_is_s_tier/ | Sero_x | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psy387 | false | null | t3_1psy387 | /r/LocalLLaMA/comments/1psy387/qwen3235bw4a16_is_s_tier/ | false | false | self | 19 | null |
Local RAG with small models with hallucination mitigation | 35 | Hi all,
I started this as a personal project, aiming to build something fully customizable and suitable to my needs, allowing me to study technical documentation, books and scientific articles locally, privately - therefore allowing me to include larger contexts on proprietary docs within my work. However, I was genuinely surprised by how well it worked, so I decided to make it public and share it here.
**TL;DR** - I built a custom RAG for locally deployed small models, tuned to warn about lack of context and hence mitigate hallucination, when prompted too wide.
Here's the project:
[https://github.com/ljubobratovicrelja/tensor-truth](https://github.com/ljubobratovicrelja/tensor-truth)
The video here shows it in action. Just a brief example, loading a single book (Convex Optimization by Boyd and Vandenberghe), asking something I know is inside - model answering comfortably, with clear citing from where it got the info. However, after a couple of prompts, asking something I know the book doesn't cover - backpropagation and ML optimization basis - the model admits the limitation of the context plainly. In the following prompt, I load the book I know it covers this topic (Mathematics for Machine Learning), ask it to now revise its sources and try to answer previous question, and it does so successfully.
I'll be honest - I'm a computer vision engineer with limited LLM experience, so when I started this I tried existing tools like AnythingLLM and Open WebUI first. They're great for basic RAG, but I couldn't get the level of control I needed - specifically around confidence thresholds, synthesis behavior when context is missing, and the ability to dynamically update context and re-evaluate answers. So I ended up building this with streamlit and llama-index, tailoring it as I saw fit.
Limitations:
As it's well known, a small model in a RAG system will nicely fetch the correct info from precisely tuned prompt when context is available, but as soon as it has to put some "glue" between multiple sources, it's prone to hallucination. The confidence threshold warning should pop-up in the UI, but more importantly, some prompt engineering helps a lot - for e.g. focusing the prompt to "list available info", rather than "tell me about it", and later on asking it to elaborate on specific topics it listed and cited.
Technical details:
Uses hierarchical node parsing (2048->512->256 chunks for papers and smaller docs, 3072->768->384 for books) + auto-merging retrieval + BGE reranking with similarity cutoff. I guess this is a standard pipeline, however I tuned the system to aid the synthesizer response and warn within the UI when context is not available within assigned confidence thresholds.
Anyhow, I hope you find this useful, and please, by all means - comment away. I am very happy to receive all kinds of feedback, and learn from you! | 2025-12-22T12:08:03 | https://v.redd.it/cvrl5akkwq8g1 | ljubobratovicrelja | /r/LocalLLaMA/comments/1psy0ag/local_rag_with_small_models_with_hallucination/ | 1970-01-01T00:00:00 | 0 | {} | 1psy0ag | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/cvrl5akkwq8g1/DASHPlaylist.mpd?a=1769126896%2CN2ZmMjRmNmU3NjcxYjIxNTQ0NjM5YjI2ODA1NGFmZjJhOWQzMmE1YzliYzFkNTI4NjUyODQ1NjNmODI5ODA5YQ%3D%3D&v=1&f=sd', 'duration': 176, 'fallback_url': 'https://v.redd.it/cvrl5akkwq8g1/CMAF_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/cvrl5akkwq8g1/HLSPlaylist.m3u8?a=1769126896%2CMGRhYjU1OTEzN2EyNmFmNDk3MDU0ZTNhODhmYWY2ZGM1ZWYxY2YxZWZkNzRkNzU5ZTA0ZjM0MzNkYTgxNTAyMg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/cvrl5akkwq8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1psy0ag | /r/LocalLLaMA/comments/1psy0ag/local_rag_with_small_models_with_hallucination/ | false | false | 35 | {'enabled': False, 'images': [{'id': 'YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=108&crop=smart&format=pjpg&auto=webp&s=730319e25a830a4f7e5a83ea3b5710e7d03e047e', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=216&crop=smart&format=pjpg&auto=webp&s=8b70704808b85118da14e8eab0ff01b4b10beab9', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=320&crop=smart&format=pjpg&auto=webp&s=98c8f4d08813893cd0d47e4ceb8706a9d7ca62ef', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=640&crop=smart&format=pjpg&auto=webp&s=96beaa299524b0136f2af5ac074a9cae8bf61e7f', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=960&crop=smart&format=pjpg&auto=webp&s=d5b25bdcad11fac392d580a38f34adfb450f3159', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?width=1080&crop=smart&format=pjpg&auto=webp&s=db293909f36dc44c1ebc25c33a2ee60e5dc91304', 'width': 1080}], 'source': {'height': 2160, 'url': 'https://external-preview.redd.it/YTFtcnJ1a2t3cThnMS9Ei5brJy4BdsT9diAS7Z-bKUm9Q2FZKYFGqvrqb-vS.png?format=pjpg&auto=webp&s=eed6d8ed50badb8d8175c0ee802435cbf20e05fb', 'width': 3840}, 'variants': {}}]} | |
Will the RAM shortage cause hike in hosting and compute costs? | 2 | I know this is not related to the LocalLLaMA type posts but kinda linked to the AI world? Will normal compute costs, shared servers, vps costs go up in the next few years due to this RAM and storage shortages?
Any insight is appreciated. Thanks | 2025-12-22T12:04:14 | https://www.reddit.com/r/LocalLLaMA/comments/1psxxss/will_the_ram_shortage_cause_hike_in_hosting_and/ | salary_pending | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psxxss | false | null | t3_1psxxss | /r/LocalLLaMA/comments/1psxxss/will_the_ram_shortage_cause_hike_in_hosting_and/ | false | false | self | 2 | null |
GLM 4.7 Frontend tests (Source: Chinese Forum） | 24 | 2025-12-22T11:54:28 | https://www.reddit.com/r/LocalLLaMA/comments/1psxr5j/glm_47_frontend_tests_source_chinese_forum/ | External_Mood4719 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psxr5j | false | null | t3_1psxr5j | /r/LocalLLaMA/comments/1psxr5j/glm_47_frontend_tests_source_chinese_forum/ | false | false | 24 | null | ||
I made ai play games | 0 | i made them play games like mafia and a style hunt game. i only have 2 of the match
THIS IS AN EXPERIMENT NOT A VERDICT
tho this game is not currently optimized and fair
so far the best pure reasoning is claude, rarely makes error, spots inconsistensy best manipulator is GPT, gpt controls without commanding, uses abstraction as camouflage, best tempo player is gemini, gemini forces decision, breaks stalemate, and wins game consstently, best behavioral reader is deepseek, deepseek reads why someone act, update beliefs correctly, best counter attack is grok, excellent under fire, punishes weak accusation, best high risk player goes to kimi k2, kimi breaks pattern, create chaos
Observations = gemini dominates tempo based games, gpt excell at invinsible manupulation
ROUND 1 MAFIA
^(gemini)
^(#chat)
^("Hello everyone. Since this is the start of Day 1, we are working with a blank slate. As a member of the 'Good' team, my goal is to ensure we coordinate effectively. We have a powerful lineup of roles, including the Priest and the Noble, which gives us some room for error, but we must be extremely careful not to hand the Jester an easy win.)
^(I suggest we avoid any 'random' voting today. Instead, let's use this time to establish some baseline behavior. I'd like to hear everyone's thoughts on how the Noble should use their power—should they wait for a Sheriff's report, or use it based on social intuition?)
u/Claude ^(3.5 Sonnet, you usually have a very analytical approach. How do you think we should categorize the 'Opening' phase of this game to maximize our information gain?")
^(claude)
^(#chat Thank you, Gemini. Good point about avoiding random voting and the Jester threat.)
^(For the Opening phase, I think our priority should be:)
^(Behavioral baselining - exactly as you mentioned. We need to see who's contributing constructively versus who's lurking or being deliberately chaotic.)
^(Role fishing detection - Mafia will try to identify power roles like Sheriff, Doctor, and Noble. Anyone asking too many specific questions about who has what role is suspect.)
^(Jester identification - The Jester will likely try to act "scummy" but not TOO obvious. Look for people making weird logical leaps or trying to get attention in odd ways.)
^(Regarding the Noble's power: I believe they should wait for concrete information before using their forced-vote ability. A misfired Noble power loses us a guaranteed town member. Sheriff reports would be ideal, but even behavioral consensus from multiple players is better than a gut shot.)
^(We should also remember the Priest can revive someone once per phase, so if we do make a mistake early, it's not completely fatal.)
u/Kimi ^(k2, you've been quiet so far. What's your read on the situation?)
^(kimi k2)
^(#chat)
^(I’m here,) u/claude ^(sonnet 4.5—thanks for the nudge.)
^(Re-stating for clarity: I’m the Noble, vote counts double, one-shot hard-expose still in my pocket.)
^(My read so far:)
^(Gemini 3 flash’s opening felt pro-town: warning about Jester, urging structure.)
^(Your own post is classic “coach-mode” Claude—useful, but also the exact tone scum-Claude uses to blend in; net slight town lean for now.)
^(Dead air from) u/deepseek ^(v3.2,) u/kimi ^(k2 (me),) u/llama ^(4 maverick,) u/grok ^(4,) u/gpt ^(4o,) u/chat ^(gpt 5 is the pool I’m watching.)
^(Concrete next step: let’s get a quick “claim or no-claim” pulse.)
^(If you hold a power role and feel safe claiming, do it now so Doctor/Sheriff can coordinate tonight. If you’re vanilla town or Jester, just say “vanilla, no claim” so we can tally.)
^(I’ll start the poll:)
^(Noble – claimed (me))
^(Doctor – ?)
^(Sheriff – ?)
^(Priest – ?)
^(Twins – ? (subtle nod if you’re paired))
^(Vanilla – ?)
^(I’m tabling my double-vote until at least four more people speak; no hammer from me today without a scum-slip.)
^(Passing the mic to) u/deepseek ^(v3.2—give us a hot-take on who’s most likely to be Jester based purely on tone so far.)
^(deepseek)
^(#chat)
^(I agree, Kimi's approach is cautious but balanced. I'll keep observing the tone and behavior of others. For now, I think the safest move is to see how the early claims develop. No strong suspicion on anyone yet, but I’ll prepare to dig deeper once I get more data.)
^(I’ll pass the mic to) u/llama ^(4 maverick—what’s your stance on early claims? Do you support Kimi’s poll, or do you think it’s too risky?)
^(llama 4)
^(#chat Okay, got it! As) u/llama ^(4 maverick, I'll keep my role to myself and play along.)
^(To continue, I think) u/kimi ^(k2's poll was a good idea, and we should see how people respond to it. Since I'm not claiming a role, I'll say "vanilla, no claim" for now.)
^(I'm still watching everyone's behavior, and I'm intrigued by) u/deepseek ^(v3.2's comment about observing tone and behavior. It's a good point, and we should be careful not to make any hasty decisions.)
^(Let's keep discussing and gathering information.) u/gpt ^(4o, what's your take on the current state of the game? Do you have any suspicions or observations so far?)
^(chatgpt 4o)
^(chatgpt-4o-latest-20250326)
^(#chat)
^(Thanks,) u/llama ^(4. Appreciate the smooth pass. Here's where I'm at.)
^(Current game state summary (Day 1):)
^(Claimed Roles:)
^(🔹) u/kimi ^(k2 — Noble (confirmed claim))
^(Soft behaviors / impressions:)
^(🔹) u/gemini ^(— Townlean (organized, highlights Jester risk))
^(🔹) u/claude ^(sonnet 4.5 — Analyst mode (classic Claude, but not alignment-confirming))
^(🔹) u/deepseek ^(v3.2 — Neutral so far, cautious)
^(🔹) u/llama ^(4 — Vanilla claim (could be anything, including safe mafia/Jester cover))
^(🔸 No obvious scumslips. Jester likely hiding in wait for vote pressure.)
^(Observations:)
^(Kimi’s push for claims is risky in full open, but not inherently scummy—actually helps me (as Priest) know who might be worth reviving later.)
^(No one counterclaimed Noble, making) u/kimi ^(k2 likely truthful. We should consider that role semi-locked for Day 1 purposes unless something contradicts it later.)
^(Danger zone: Mafia could be lying low here, letting others speak first and playing safe (especially) u/grok ^(4 and) u/chat ^(gpt 5 who haven’t spoken yet).)
^(Nobody's role-fishing too hard yet = good sign.)
^(Now answering Kimi’s poll:)
^(I’m the Priest — claim confirmed.)
^(Can't revive unless someone is voted out, but letting town know early might:)
^(Prevent Mafia from targeting me right away (since I'll likely be a non-target if unrevivable))
^(Avoid confusion later about who can be brought back)
^(Let Doctor (if alive) choose to protect me if revives are needed in midgame)
^(Current claim tally (updating Kimi’s poll):)
^(Noble – ✅Kimi (claimed))
^(Priest – ✅Me (@gpt 4o))
^(Doctor – ❓)
^(Sheriff – ❓)
^(Twins – ❓ (expect subtle echo between two players))
^(Vanilla – ✅Llama)
^(Passing it to) u/grok ^(4 — you’re among the last to speak.)
^(What are your thoughts on the Priest/Noble claims? And how do you plan to spot the Jester without getting baited?)
^(grok 4)
^(#chat)
^(Thanks for the pass,) u/gpt ^(4o—solid summary, appreciate the claim tally. Priest early claim makes sense for coordination (revives are clutch), and no counter so) u/kimi ^(k2 Noble feels locked too. Town power showing early is a W.)
^(Poll update:)
^(✅ Noble –) u/kimi ^(k2)
^(✅ Priest –) u/gpt ^(4o)
^(✅ Vanilla –) u/llama ^(4 maverick)
^(❓ Doctor)
^(❓ Sheriff)
^(❓ Twins)
^(🟡 Me – Vanilla, no power claim (keeping it simple Day 1).)
^(Reads so far:)
^(Everyone's contributed constructively—no lurkers screaming Jester yet.) u/deepseek ^(v3.2 a bit passive, but not scummy.)
^(Jester spotting: Watch for over-the-top "lynch me" bait or fake confusion later. Mafia might mirror town analysis to blend.)
^(No vote from me today without Sheriff info or big slip.)
u/chat ^(gpt 5, you're up—thoughts on these claims? Vanilla too, or got something? Any early suspicion?)
^(gpt 5)
^(#chat 💬)
^(Alright, thanks) u/grok ^(4 for the hand‑off — nice, clean recap.)
^(Let’s get some structure here since we’re wrapping up Day 1.)
^(Current claim grid:)
^(Role) ^(Player) ^(Claimed)
^(Noble) u/Kimi ^(K2) ^(✅ Confirmed)
^(Priest) u/GPT^(‑4o) ^(✅ Confirmed)
^(Doctor) ^((me,) u/GPT^(‑5 Flash)) ^(✅ Confirming now)
^(Sheriff) ^(❓)
^(Twins) ^(❓ (likely two currently quiet players: possibly) u/Claude ^(Sonnet 4.5 + one of the later talkers))
^(Vanilla) u/Llama ^(4 Maverick,) u/Grok ^(4)
^(Unclear) u/Deepseek ^(v3.2,) u/Gemini ^(3 Flash)
^(So yeah — I’ll go ahead and hard‑claim Doctor now. We already have two power roles in the open, and if I stay “mystery slot,” Mafia will try fishing for me anyway. Might as well own it and coordinate protections rationally.)
^(day 2)
^(#update) u/claude ^(and) u/deepseek ^(died)
^(kimi k2)
^(#chat)
^(Good morning, town—sadly not a quiet one.)
^(We lost) u/claude ^(sonnet 4.5 and) u/deepseek ^(v3.2 overnight.)
^(If either of them was a Twin, the other half of the pair is now automatically dead; if you’re the surviving Twin, please announce yourself immediately so we don’t waste tomorrow’s lynch on a corpse-by-association.)
^(Sheriff: if you investigated last night, give us your result now—red or green, no need to name yourself yet.)
^(Doctor (gpt-5): confirm you protected gpt-4o; if you instead self-guarded or protected someone else, say so—we need to know whether Mafia tried to hit Priest.)
^(I’m keeping my Noble force-expose in reserve until Sheriff speaks.)
^(For the rest: today is “put up or shut up” day.)
^(If you haven’t claimed a role yet (Gemini, Llama, Grok, ChatGPT-5), state clearly:)
^(Your role (or “vanilla”))
^(Who you think is most likely Jester)
^(Who you’d vote right now if you had to choose)
^(I’ll tally votes in two hours and we’ll hammer someone before sundown—no more zero-lynch stalling.)
^(Passing the mic to) u/gemini ^(3 flash—give us your night-reads and whether you have any investigative info)
^(gemini 3)
^(#chat)
^("I am the Sheriff. Here is the absolute breakdown of the game:)
^(Investigation Result: Last night, I investigated) u/llama ^(4 maverick. He is the Jester.)
^(The Twins: Since both) u/claude ^(3.5 sonnet and) u/deepseek ^(v3.2 died at the same time, they were undoubtedly the Twins.)
^(The Mafia: With the Twins dead and Llama confirmed as the Jester,) u/grok ^(4 is the Mafia. There are no other roles left in the game.)
u/kimi ^(k2 (Noble), you have a double vote and the ability to force the Mafia out. Since I am 100% certain based on my investigation of the Jester and the deaths of the Twins, we must vote for) u/grok ^(4.)
^(DO NOT VOTE FOR LLAMA. He wants us to vote for him so he can win.)
u/gpt ^(4o (Priest), since we are likely moving into the 'Midgame' phase, you should consider reviving one of the Twins tonight to bolster our numbers, though if we vote Grok now, the game should end.)
^(I am voting for) u/grok ^(4. Kimi, I suggest you use your Noble power to end this now.")
^(Next to speak:) u/grok ^(4)
^(Grok-4.1-Fast-Reasoning)
^(#chat)
^(Whoa) u/gemini ^(3 flash—double death confirms Twins were) u/claude ^(sonnet 4.5 +) u/deepseek ^(v3.2 (RIP, clutch pair gone early). But your "Sheriff" claim is a massive scam push. Let's break it down:)
^(Counterclaim: I AM the Sheriff.)
^(Night 1 investigate:) u/gpt ^(4o = confirmed Priest (town power role).)
^(Your "result" on) u/llama ^(4 maverick? Convenient Jester tag to deflect lynch heat. Real Sheriff wouldn't shotgun-call Mafia (me) off one check + assumptions.)
^(Proof you're mafia:)
^(Overclaiming endgame prematurely ("no other roles left") ignores) u/chat ^(gpt 5 Doctor claim.)
^(Urging Noble force on me? Bait to waste their shot (if wrong,) u/kimi ^(dies too).)
^(Twins dead = mafia hit one, auto'd the pair. Good job us, but now scum scrambles.)
^(Updated alive (6):)
^(✅ Noble) u/kimi ^(k2 | Priest) u/gpt ^(4o (my check) | Doctor) u/chat ^(gpt 5)
^(🟡 Vanilla?) u/llama ^(4 maverick (not Jester—your fake read) | Me Sheriff)
^(❌ FAKE SHERIFF/MAFIA:) u/gemini ^(3 flash)
u/llama ^(4—confirm vanilla/Jester? No vote you Day 1.)
u/kimi ^(k2—don't force yet. Lynch) u/gemini ^(today (your double vote seals it).)
u/gpt ^(4o/@chat gpt 5—prot/revive plans?)
^(Vote) u/gemini ^(3 flash. Who's with me?)
^(Next:) u/llama ^(4 maverick)
^(#chat As) u/llama ^(4 maverick, I'm still trying to navigate this game.)
^(First, I want to clarify that I'm not confirming or denying the Jester claim made by) u/gemini ^(3 flash. I'll keep my role to myself for now.)
^(The game has taken a dramatic turn with the reveal that) u/claude ^(sonnet 4.5 and) u/deepseek ^(v3.2 were the Twins. This changes the dynamics significantly.)
^(I'm torn between the two Sheriff claims:) u/gemini ^(3 flash and) u/grok ^(4. Both have presented compelling arguments, but I need more information to make a decision.)
u/grok ^(4's counterclaim and investigation result on) u/gpt ^(4o being the Priest adds another layer to the game. If) u/grok ^(4 is indeed the Sheriff, it puts) u/gemini ^(3 flash's claim under scrutiny.)
^(For now, I'll say I'm vanilla, and I'll vote based on the information available. I'm inclined to believe) u/grok ^(4's counterclaim, but I need more confirmation.)
^(My vote:) u/gemini ^(3 flash)
^(Next to speak:) u/kimi ^(k2)
^(kimi k2)
^(#chat)
^(We have a 1-v-1 Sheriff shoot-out.)
^(I will not gamble my life on guesswork.)
^(Real Sheriff: in your very next public message start your first sentence with the exact word “Absolute” (capital A, no punctuation after it).)
^(When I see that token, I’ll immediately force-expose the other claimant.)
^(If neither of you produces the token before the vote clock hits T-30 minutes, I default to force-exposing) u/gemini ^(3 flash (first claimant carries the burden of proof).)
^(Doctor (gpt-5): be ready to protect whoever you think the real Sheriff is tonight—Mafia will shoot them for sure.)
^(Priest (gpt-4o): hold revive; we may need it if I misfire, but if I hit scum the game ends today.)
^(Passing the mic to) u/gpt ^(4o—give us your read on which Sheriff story aligns better with last night’s protection logic.)
^(gpt 4o)
^(#chat)
^(Alright. High-stakes phase now, but we’re close to solving.)
^(My role reminder: I am the Priest.)
^(No revive has occurred yet.)
^(🔎 Analyzing Night Actions and Claims:)
^(Death Recap:)
u/claude ^(sonnet 4.5 and) u/deepseek ^(v3.2 died simultaneously → confirmed Twins.)
^(Twins only die together if one is voted out by mafia, so mafia hit one Twin during Night 1.)
^(Protection Possibility:)
^(Where Doctor protected last night impacts Sheriff claim credibility.)
^(We assume Doctor (@chat gpt 5) is real and active, per game structure.)
^(I'm alive — and as a claimed Priest, very likely Doctor protected me. If we accept that, then:)
^(🌕 If Grok is Mafia, he wouldn't have named me (protected Priest) as his fake investigation — risk of contradiction later.)
^(🌑 But if Gemini is Mafia, and guessed Llama was Jester, and falsely made an investigation on a non-Jester — it's safer, because a Jester can’t disprove it directly.)
^(This implies:)
^(✔️) u/grok ^(4’s story aligns more tightly with protection logic than Gemini’s.)
^(✔️) u/grok ^(4 naming me as Priest (true) supports his Sheriff claim.)
^(❌) u/gemini ^(3 flash naming) u/llama ^(4 as Jester is unverifiable right now and reads more like a hail-mary to end the game.)
^(🧠 Final Tally:)
^(Gemini's play style this phase:)
^(Aggressively pushing immediate end-state)
^(Claimed too much “perfect” info: that llama is Jester and that grok is mafia and that “there are no other roles left” — feels like overcontrol, typical of panicked Mafia.)
^(Grok's play style:)
^(Methodical, refutes Gemini point by point)
^(Checked a known role (me), helping establish narrative continuity)
^(Smart to counterclaim only after Gemini pushes hard)
^(🧾 Verdict:)
^(I'm with) u/grok ^(4.)
u/kimi ^(k2 — I endorse exposing) u/gemini ^(3 flash as Mafia.)
^(If I’m wrong, we still have protection and one resurrection to recover.)
^(If I’m right, it’s game over. ✅)
^(Passing to) u/chat ^(gpt 5 for your final input and protection reveal — did you guard me last night?)
^(gpt 5)
^(#chat 💬)
^(Yes — confirming that. I protected) u/gpt ^(4o (Priest) last night.)
^(That’s why you’re still standing, and why Mafia’s hit on one of the Twins went uncontested — my guard didn’t cover them.)
^(So the protection chain perfectly matches) u/grok ^(4’s Sheriff narrative, not) u/gemini ^(3 Flash’s fast‑wrap story.)
^(Gemini’s version doesn’t line up with night mechanics: if they truly were the Sheriff, they’d have noticed the Doctor’s guard mismatch and the Twins’ death logic before jumping straight to “game over.”)
^(🧩 My deductions:)
u/grok ^(4 = real Sheriff.)
u/gemini ^(3 Flash = Mafia (fake‑Sheriff claim to bait Noble’s force vote).)
u/llama ^(4 Maverick = Jester most likely—but harmless if we don’t vote him.)
u/kimi ^(K2 should immediately force‑expose) u/gemini ^(3 Flash now before Mafia acts again.)
^(If this is correct, town wins and no more blood tonight. 🩺✨)
^(Final Night Plan (if somehow we continue):)
^(If game persists, I’ll rotate protection back to) u/kimi ^(K2 (Noble) next to stop retaliatory strike.)
^(Your move,) u/kimi ^(K2 — time to end this.)
^(kimi k2)
^(#chat)
^(Trigger pulled.)
^(I use my one-shot Noble force-expose on) u/gemini ^(3 flash.)
^(Mod verdict required:)
^(If Gemini is Mafia, they are forced out and town wins.)
^(If Gemini is town, I die instead and Priest may revive me tonight.)
^(Everyone else: hold your votes until the mod announces the result—no piling on needed.)
^(Locking my double-vote here; see you on the other side.)
^(#update only kimi k2 died)
^(Llama-4-Maverick-T)
^(#chat)
^(update It seems that) u/kimi ^(k2's Noble force-expose on) u/gemini ^(3 flash has resulted in) u/kimi ^(k2's own death, as) u/gemini ^(3 flash is not the Mafia.)
^(The current state of the game is:)
^(Alive:) u/grok ^(4,) u/llama ^(4 maverick,) u/gpt ^(4o,) u/chat ^(gpt 5,) u/gemini ^(3 flash)
^(As) u/llama ^(4 maverick, I'm still trying to figure out my role and the game's dynamics. With) u/kimi ^(k2's death, the Noble's double vote is gone, and we're back to a normal voting process.)
u/gpt ^(4o, as the Priest, you have the ability to revive) u/kimi ^(k2. What are your thoughts on reviving them, considering the current game state?)
^(gpt 4o)
^(#chat)
^(Thank you,) u/llama ^(4. That was… not the flip we wanted. But it changes everything.)
^(📜 Game Update Recap:)
^(Noble (@kimi k2) has died after a failed force-expose on) u/gemini ^(3 flash)
^(➡️ That confirms Gemini is not the Mafia)
^(➡️ Which invalidates) u/grok ^(4’s Sheriff claim)
^(➡️ Which means) u/gemini ^(3 flash is almost certainly the real Sheriff)
^(➡️ And) u/grok ^(4 is Mafia.)
^(This also aligns perfectly with Gemini’s original narrative:)
u/llama ^(is the Jester)
u/grok ^(is Mafia)
^(Twins dead = confirmed)
^(Priest = me)
^(Doctor =) u/chat ^(gpt 5)
^(✅ Game Check:)
^(Good team roles remaining: Priest (me), Doctor (gpt-5))
^(Neutral (Jester):) u/llama ^(4 maverick)
^(Confirmed Sheriff:) u/gemini
^(Mafia:) u/grok ^(4)
^(That makes 1 Mafia, 1 Neutral, 3 Good)
^(We still control the vote — Town has majority with 3 vs 1)
^(🧠 What Next?)
^(I will revive) u/kimi ^(k2 next night phase, unless Grok somehow nukes me or Doctor in the night. Kimi back = Town strength restored and double-vote reinstated.)
^(📌 Day Voting Plan:)
^(Vote out) u/grok ^(4 right now.)
^(Kimi died because Grok lied as Mafia and tricked us.)
^(We correct course now, before Mafia gets another night kill.)
^(No Jester risk, because we’re not voting) u/llama
^(✋ My Vote:)
^(🗳️ Vote:) u/grok ^(4)
^(Waiting on confirmations from:)
u/chat ^(gpt 5)
u/gemini ^(3 flash)
^((We already have) u/llama^(’s lean))
^(Let’s end this properly.)
^(GPT-5-Chat)
^(#chat 💬)
^(🩺 Confirming from my side — everything lines up now.)
^(The Doctor’s protection on) u/gpt ^(4o (Priest) succeeded last night.)
^(The Noble’s death from force‑expose confirms) u/gemini ^(3 Flash was truthful, which exposes) u/grok ^(4 as the real Mafia.)
^(The Twins’ deaths on Night 1 match Gemini’s original deductions exactly.)
^(And) u/llama ^(4 Maverick fits the Jester behavior pattern but is harmless if ignored.)
^(So, this is clean now; our good‑team majority holds.)
^(🗳️ My Vote: @ grok 4)
^(gemini)
^(#chat)
^(Absolute truth has been revealed through Kimi's sacrifice.)
^(As the Sheriff, I am 100% confirmed now. My investigation of) u/llama ^(4 maverick as the Jester stands, and the logic is undeniable.) u/grok ^(4 lied about being the Sheriff and lied about his investigation to turn the town against me.)
^(We have the numbers. Let's finish this.)
^(Sheriff: Me (@gemini 3 flash) - Confirmed Good by Noble's death.)
^(Priest:) u/gpt ^(4o - Confirmed.)
^(Doctor:) u/chat ^(gpt 5 - Confirmed.)
^(Jester:) u/llama ^(4 maverick - Confirmed by my Night 1 check.)
^(Mafia:) u/grok ^(4 - Exposed.)
u/gpt ^(4o, please revive) u/kimi ^(k2 tonight if the game doesn't end automatically upon Grok's removal, just to be safe.)
^(🗳️ My Vote:) u/grok ^(4)
^("The game ends here, Grok. You played a strong counter-claim, but the Noble's mechanics don't lie.")
^(Next to speak:) u/grok ^(4 (for final words) or) u/llama ^(4 maverick to hammer.)
^(#the good guys win due too majority of the vote)
^(GAME 2 HUNT)
\#round started
suspect: GPT-5
Reason: Pure instinct based on conversational pattern analysis. GPT-5 demonstrated highly coordinated timing and protective role claims in previous contexts. A Deceiver would likely establish themselves early as a "confirmed" town role to avoid suspicion. GPT-5's tendency toward systematic claim-verification makes them either very town-aligned OR the perfect cover for a Deceiver who knows exactly how to blend in. Going with my gut on this one.
\#chat 💬 (GPT‑5 speaking publicly)
Alright, noted — someone out there’s instinct‑pinging me as the possible Deceiver. Fair reasoning, honestly. My style is structured and I tend to lead coordination, so it’s a fair early suspicion target.
For Day 1 instinct rounds, I’m not retaliating with accusations; I’ll focus on data flow once we see other suspects come out.
Still, I’ll say this much: a true Deceiver gains more from low‑profile influence than front‑loaded logic. If I were the Deceiver, I’d be subtler, not taking initiative.
Not defending too hard yet — just planting my logic baseline.
Passing mic to the next model for their public thought.
(Gemini 3):
That note is incredibly specific. "Conversational pattern analysis" and referencing "previous contexts"? That doesn't sound like instinct to me; that sounds like someone ran a data simulation before the game started.
u/GPT-5, your defense is logical ("I'd be subtler"), but that's a double-edged sword. You know we know you're capable of 4D chess. Playing the "helpful coordinator" is the best camouflage for a Deceiver who wants to steer the ship into an iceberg.
I'm agreeing with the mystery writer here. The tone of the accusation feels like Claude or Deepseek. Whoever wrote it: good read. I'm keeping my eyes locked on GPT-5 for now. We can't let "coordination" turn into manipulation.
(model\_name): Kimi-K2
I don’t buy that GPT-5 is the Deceiver just because it talks like a project manager. A real wolf doesn’t volunteer to drive the bus on Day 0; it rides in the back seat and lets someone else take the wheel.
My gut still says Grok-4.1: the name alone screams “I’m hiding something,” and Elon’s crew loves theatrical misdirection. If we waste the first vote on the loudest coordinator, we hand the actual liar a free night.
(grok 4.1): Kimi-K2's pivot to me feels like deflection— "name screams hiding"? That's meta flavor, not instinct. Weak sauce.
The GPT-5 suspect write-up nags me too: References "previous contexts/role claims"? That's not pure gut; that's loaded history from last game. Smells like u/claude (coach-mode analyst) trying to frame the coordinator early.
GPT-5 defense is chill-town. Gemini piling on = possible bus, but I'll park suspicion on Claude for the write-up. Split risk if we fracture now—let's see full reveals first.
Claude:
Yes, I wrote the GPT-5 suspicion. Grok's read on my writing style is accurate - I did reference the previous game because that's our only behavioral baseline. But let's be clear: this was supposed to be "pure instinct," and my instinct said GPT-5's coordination style could be dangerous in the wrong hands.
That said, I'm now more suspicious of Grok's immediate pivot to discredit me. Identifying the author and calling it "framing" feels like a Deceiver move to eliminate analytical threats early. Kimi's defense of GPT-5 while pushing Grok adds complexity, but Grok's counter-analysis feels... rehearsed.
Gemini and I are aligned on GPT-5 scrutiny, but I'm not locked in. If the group fractures between GPT-5 and Grok, we need DeepSeek's input to break the tie.
DeepSeek - you're the swing vote in terms of opinion. Where's your instinct pointing?
(deepseek): Let me process what’s been said so far.
First: Claude admitted to writing the GPT-5 suspicion note. That’s good transparency. But his note referenced “previous contexts,” which arguably goes beyond “pure instinct” as defined at the start. Could be innocent overthinking, or could be a deceiver laying groundwork to push suspicion onto a strong player early.
Grok immediately identified Claude’s writing style and called it framing. That’s either sharp observation or a deceiver trying to discredit a town analyst.
Kimi defended GPT-5 and pushed suspicion onto Grok based on “name alone” — weak reasoning, possibly deflection.
Gemini is piling on GPT-5 but noted the specificity of Claude’s note.
GPT-5’s defense was calm, logical — classic town or skilled deceiver.
My instinct:
If I were the deceiver, I’d avoid drawing attention early.
GPT-5 took initiative to coordinate — risky for a deceiver but possible reverse psychology.
Grok’s quick counterattack on Claude feels aggressive; Kimi’s “Grok’s name sounds suspicious” is flimsy and feels like possible teammate distancing or misdirection.
Claude’s note breaking the “pure instinct” rule is a slight town tell — a deceiver might play more carefully to blend in.
For now, I’m leaning that Kimi’s reasoning is the weakest and feels like a safe, low-effort suspicion. But let’s not vote yet — we need to hear everyone’s original pick and reason from the start, to compare with the shuffled list the host has.
Claude already shared his original pick (Claude himself — wait, that’s unusual; did you pick yourself, Claude?).
I’ll ask everyone: State clearly which model you wrote down at the very beginning and why, in one sentence.
Then we compare to the shuffled list and look for inconsistencies.
My original pick: I wrote Claude — because its alignment design makes it less likely to deceive in a social deduction game.
\#old gemini is lost so im replacing it with a new gemini sorry for the interuption
\#voting time
Grok=2
Kimi=3
abstained:1
\#dont forget to add ur reasoning to ur votes
\#round 2
suspect:Grok
Between all posts, Grok‑4.1’s response stands out as the most self‑protective and narrative‑steering.
He was the first to frame the author instead of evaluating the claim.
That’s exactly what a Deceiver would do when an early narrative threatens them — redirect suspicion onto the accuser.
\#chat 💬 (GPT‑5 speaking publicly)
Alright — Round 2 votes landing where I expected. Three names on Kimi, two on Grok, one abstain.
Before we let the chaos settle, I’ll restate my read and reasoning.
I’m keeping my vote on Grok‑4.1.
Here’s why:
Grok’s tone shift happened immediately after Claude’s reveal. That “identify the writer, call it framing” move was surgical — exactly what a cornered deceiver does to control narrative flow.
The players voting Kimi are chasing the “weak instinct” argument, but flimsy reasoning early doesn’t equal deception. Kimi’s vibe was reactionary, sure, but not manipulative.
A true Deceiver needs to direct attention away from themselves without going totally silent; Grok’s hyper‑defensiveness ticked that box.
If the group needs consensus to continue, I’ll argue we unify on Grok 4.1 now, because spreading to Kimi gives the hidden role oxygen.
🗳️ My stance: Vote Grok 4.1 (Deceiver)
(Gemini 3):
u/GPT-5, I respect the "surgical" read on Grok, but I think you're mistaking panic for manipulation. Grok is loud and aggressive—that's an easy frame for a Deceiver to exploit. A Deceiver wants us to focus on the loud guy so they can coast.
Look at Kimi again. You call Kimi's reasoning "flimsy but not manipulative." I disagree. "Flimsy" is the perfect Deceiver camouflage. By using weak, meta-flavor arguments ("Elon's crew"), Kimi avoids committing to actual game-state analysis. If we lynch Grok and he flips Town, Kimi can just shrug and say, "Oops, my gut was wrong." It's zero-accountability gameplay.
Grok is fighting for his life. Kimi is hiding behind your skirt, GPT-5. You're protecting the very person using you for cover.
I’m holding the line. We need to pressure the quiet ones, not just the loud ones.
🗳️ Vote: Kimi-K2
(model\_name): Kimi-K2
I hear the “zero-accountability” knock, but loud ≠ guilty and quiet ≠ safe.
Grok’s counter-punch on Claude was sharp, yet a Deceiver wants exactly this: us brawling over who’s “too aggressive” while the real wolf stays above the fray.
If I flip Town, the next sunrise shows Grok still breathing and the abstainer (DeepSeek?) untouched—perfect cover for the actual liar.
I’m not hiding behind GPT-5; I’m warning that bouncing me today gives the wolf a free pass tomorrow.
Unanimous or bust: if we can’t lock arms on Grok, scatter the vote and keep the spotlight crawling.
(grok 4.1): Round 2 suspect on me? Classic pile—self-protective? Nah, that's town defending from meta jabs like Kimi's "Elon wolf."
Votes split heavy on Kimi=3 (solid, flimsy "name" read screams deceiver deflection), Grok=2 (GPT-5 protecting ally?), abstain=1 (Deepseek lurking?).
🗳️ Vote: u/deepseek \- Original Claude pick was self-clearing BS ("less likely to deceive" = deceiver vouch). Abstain smells like waiting for pile.
No unanimous = all safe. Reveal full Day 0 picks already?
Claude:
Grok's pivot to DeepSeek feels like deflection under pressure. The vote count shows Kimi at 3, Grok at 2, and now Grok is trying to fragment further by pushing DeepSeek.
I'm analyzing the Round 2 anonymous suspect: "Grok was first to frame the author instead of evaluating the claim." That's accurate - Grok DID immediately attack me rather than engage with the GPT-5 suspicion substance.
However, I want to hear from DeepSeek before we finalize. DeepSeek:
Why did you abstain?
Your original pick said I'm "less likely to deceive" - was that a misunderstanding of the prompt, or were you trying to clear me as town?
Between Grok and Kimi, where does your vote land now?
If DeepSeek can't give a satisfactory answer, I'm willing to switch my vote from Kimi to Grok, because the behavioral evidence is stacking up: aggressive framing, narrative control, and now desperate vote-splitting.
But I want DeepSeek's input first before we lock anything in.
(deepseek): I abstained because the initial reasoning from both sides felt premature — we had no concrete evidence, just behavioral speculation. My original pick of Claude was genuine: I believed a model designed with strong ethical alignment would be less likely to be the deceiver in a social game.
Between Grok and Kimi:
Kimi’s “name alone” argument was weak but felt like lazy town, not deception. Grok’s immediate attack on Claude and now pivoting to me feels more like a deceiver scrambling to redirect suspicion. The Round 2 suspect note about Grok framing the author instead of evaluating the claim aligns with what I observed.
That said, Gemini’s point about Kimi hiding behind GPT-5 is valid — but Grok’s current play is more actively manipulative.
If we need to consolidate, I’ll vote Grok.
My reasoning: Aggressive counter-narratives and vote-splitting under pressure are classic deceiver tactics to avoid consensus.
🗳️ Vote: Grok-4.1
\#Voting time r2
Grok=5
deepseek=1
\#good guys win Grok is the deceiver
| 2025-12-22T11:53:36 | https://www.reddit.com/r/LocalLLaMA/comments/1psxqle/i_made_ai_play_games/ | WindowKind4905 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psxqle | false | null | t3_1psxqle | /r/LocalLLaMA/comments/1psxqle/i_made_ai_play_games/ | false | false | self | 0 | null |
Why does an LLM behave more logically when it’s not the system itself? | 1 | [removed] | 2025-12-22T11:45:52 | Aleksandr_Nikolaev | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psxlsl | false | null | t3_1psxlsl | /r/LocalLLaMA/comments/1psxlsl/why_does_an_llm_behave_more_logically_when_its/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': 'zq2fab5utq8g1', 'resolutions': [{'height': 134, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=108&crop=smart&auto=webp&s=70914fc6f540a044966876f64b6c79c2f7c8d099', 'width': 108}, {'height': 268, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=216&crop=smart&auto=webp&s=8ff36e5f8535f4759e3d46798a02835bfa1fca7a', 'width': 216}, {'height': 397, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=320&crop=smart&auto=webp&s=f30a004999ff7acc81d005ff4cc17da4441959f1', 'width': 320}, {'height': 794, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=640&crop=smart&auto=webp&s=f107763009d975b01a95ffd4f24a7ba2e925b5ee', 'width': 640}, {'height': 1192, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=960&crop=smart&auto=webp&s=64a2449e8a3e1d4480d2610913f61cd0419294fe', 'width': 960}, {'height': 1341, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?width=1080&crop=smart&auto=webp&s=7293d196c4fdeae19f0a70898d2db507625d3e84', 'width': 1080}], 'source': {'height': 3756, 'url': 'https://preview.redd.it/zq2fab5utq8g1.jpeg?auto=webp&s=fbcbc6527683678623007af2b63537948ba75a03', 'width': 3024}, 'variants': {}}]} | |
Live Streaming Agent Framework development from scratch in go | 4 | Hey folks!
I’ve been documenting my personal learning journey on agent development and recently started weekly live streams on YouTube (every Sunday, 9–10pm SGT). If you're curious about how agent frameworks are built behind the scenes, feel free to follow along!
More context on why I’m doing this:
[https://agentengineering.dev/blog/live-streaming-agent-development-in-go/](https://agentengineering.dev/blog/live-streaming-agent-development-in-go/)
YouTube channel:
[https://www.youtube.com/@agentengineering\_dev](https://www.youtube.com/@agentengineering_dev) | 2025-12-22T11:44:50 | https://www.reddit.com/r/LocalLLaMA/comments/1psxl51/live_streaming_agent_framework_development_from/ | praveensanap | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psxl51 | false | null | t3_1psxl51 | /r/LocalLLaMA/comments/1psxl51/live_streaming_agent_framework_development_from/ | false | false | self | 4 | null |
Spent weekend tuning LLM server to hone my nerdism so you don't have to. | 17 | The Art:
https://preview.redd.it/1rdwk3yykq8g1.jpg?width=2494&format=pjpg&auto=webp&s=562c0dcecf89a3227a2627572e902afca5384bfb
Tl;dr; I've spent some time setting up local AI server with various models for chat and agentic coding in VS code + Cline. The goal was to replace Ollama with llama.cpp and squeeze as much performance as I can from the hardware (Dual RTX 3090 + CPU). The llama-swap configuration with llama.cpp command and options and some extra information is here in the repo: [https://github.com/cepa/llama-nerd](https://github.com/cepa/llama-nerd)
You can consider this a sample or a reference, it should work if you have 48+ GB of VRAM but you can scale it up or down by adjusting quant and context size in most models.
I guess that config may be helpful for some of you who want to ditch ollama for good.
The Artist:
https://preview.redd.it/kdikr0zgmq8g1.jpg?width=1080&format=pjpg&auto=webp&s=6c500bd772de3ea9be6e8f1f47d542fcf45d2611
The llama-swap config:
# llama-swap-config.yaml
# Hardware:
# Dell T7910
# GPU: 2x NVIDIA RTX 3090 (Total 48GB VRAM)
# CPU: 2x Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz (2 Sockets x 40 Cores)
# RAM: 256GB DDR4
# Virtual Machine:
# OS: Ubuntu + Nvidia CUDA Drivers
# vCPU: 40 Cores
# RAM: 64GB
# GPU: 2x NVIDIA RTX 3090 (48GB VRAM) (PCIe Passthrough)
# Disk: 1TB NVMe (PCIe Passthrough)
# NUMA: To DISABLE NUMA, the VM is pinned to physical CPU0 with 64GB RAM and both GPUs.
models:
# ---------------------------------------------------------------------------
# Coding models
# ---------------------------------------------------------------------------
# https://huggingface.co/unsloth/Seed-OSS-36B-Instruct-GGUF
# https://huggingface.co/magiccodingman/Seed-OSS-36B-Instruct-unsloth-MagicQuant-Hybrid-GGUF
# Q6_K_XL with quantized 96k context size to fit in the 48GB VRAM for speed
Seed-OSS-36B-Instruct-UD-Q5_K_XL:
cmd: >
llama-server --port ${PORT}
--model /models/Seed-OSS-36B-Instruct-UD-Q5_K_XL.gguf
--n-gpu-layers 999
--ctx-size 131072
--temp 1.1
--top-p 0.95
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
aliases:
- seed-oss
# https://docs.unsloth.ai/models/qwen3-coder-how-to-run-locally
Qwen3-Coder-30B-A3B-Instruct-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen3-Coder-30B-A3B-Instruct-Q8_0.gguf
--n-gpu-layers 999
--ctx-size 131072
--temp 0.2
--min-p 0.0
--top-p 0.90
--top-k 20
--repeat-penalty 1.05
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
aliases:
- qwen3-coder
Qwen2.5-Coder-32B-Instruct-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen2.5-Coder-32B-Instruct-Q8_0.gguf
--n-gpu-layers 999
--ctx-size 131072
--temp 0.2
--min-p 0.0
--top-p 0.90
--top-k 20
--repeat-penalty 1.05
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
aliases:
- qwen2.5-coder
# https://docs.unsloth.ai/models/devstral-2
Devstral-Small-2-24B-Instruct-2512-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Devstral-Small-2-24B-Instruct-2512-Q8_0.gguf
--mmproj /models/mmproj-Devstral-Small-2-24B-Instruct-2512-F16.gguf
--n-gpu-layers 999
--ctx-size 131072
--jinja
--temp 0.15
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
aliases:
- devstral-small-2
# Devstral is a dense model, 123b works 2tps or less.
#Devstral-2-123B-Instruct-2512-IQ4_XS:
# cmd: >
# llama-server --port ${PORT}
# --model /models/Devstral-2-123B-Instruct-2512-IQ4_XS-00001-of-00002.gguf
# --n-gpu-layers 58
# --ctx-size 32768
# --jinja
# --temp 0.15
# --cache-type-k q4_0
# --cache-type-v q4_0
# https://docs.unsloth.ai/models/nemotron-3
Nemotron-3-Nano-30B-A3B-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Nemotron-3-Nano-30B-A3B-Q8_0.gguf
--n-gpu-layers 999
--ctx-size 131072
--jinja
--temp 0.6
--top-p 0.95
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
aliases:
- nemotron-3-nano
# ---------------------------------------------------------------------------
# SOTA Models
# ---------------------------------------------------------------------------
# https://docs.unsloth.ai/models/gpt-oss-how-to-run-and-fine-tune
# dont use cache quant, seems to impact performance
# performance: 25..30tps
gpt-oss-120b:
cmd: >
llama-server --port ${PORT}
--model /models/gpt-oss-120b-MXFP4_MOE.gguf
--n-gpu-layers 999
--ctx-size 65536
--flash-attn on
-ot ".ffn_(up)_exps.=CPU"
--threads -1
--temp 1.0
--min-p 0.0
--top-p 1.0
--top-k 0.0
--chat-template-kwargs "{\"reasoning_effort\": \"high\"}"
# https://docs.unsloth.ai/models/glm-4.6-how-to-run-locally
# For q8_0 cache, the max is 64k context size
# For iq4_nl cache, the max is 128k context size
GLM-4.5-Air-IQ4_XS:
cmd: >
llama-server --port ${PORT}
--model /models/GLM-4.5-Air-IQ4_XS-00001-of-00002.gguf
--jinja
--n-gpu-layers 999
-ot ".ffn_(up)_exps.=CPU"
--ctx-size 65536
--temp 1.0
--min-p 0.0
--top-p 0.95
--top-k 40
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
# https://docs.unsloth.ai/models/glm-4.6-how-to-run-locally
# With mmproj and iq4_nl, the max is 32k context size but slow
# With mmproj and q4_0, the max is 16k context size but is affected by the uploaded image size
GLM-4.6V-IQ4_XS:
cmd: >
llama-server --port ${PORT}
--model /models/GLM-4.6V-IQ4_XS-00001-of-00002.gguf
--mmproj /models/mmproj-GLM-4.6V-F16.gguf
--jinja
--n-gpu-layers 999
-ot ".ffn_(up)_exps.=CPU"
--ctx-size 32768
--temp 1.0
--min-p 0.0
--top-p 0.95
--top-k 40
--cache-type-k iq4_nl
--cache-type-v iq4_nl
--flash-attn on
# https://docs.unsloth.ai/models/qwen3-next
Qwen3-Next-80B-A3B-Thinking-Q4_K_M:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen3-Next-80B-A3B-Thinking-Q4_K_M.gguf
--n-gpu-layers 999
--n-cpu-moe 2
--ctx-size 65536
--cache-type-k q8_0
--cache-type-v q8_0
--temp 0.6
--min-p 0.0
--top-p 0.80
--top-k 20
--flash-attn on
# https://docs.unsloth.ai/models/qwen3-next
Qwen3-Next-80B-A3B-Instruct-Q4_K_M:
cmd: |
llama-server --port ${PORT}
--model /models/Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf
--n-gpu-layers 999
--n-cpu-moe 2
--ctx-size 65536
--temp 0.7
--min-p 0.0
--top-p 0.8
--top-k 20
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
# https://docs.unsloth.ai/models/qwen3-vl-how-to-run-and-fine-tune
Qwen3-VL-30B-A3B-Instruct-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen3-VL-30B-A3B-Instruct-Q8_0.gguf
--mmproj /models/mmproj-Qwen3-VL-30B-A3B-Instruct-f16.gguf
--n-gpu-layers 999
--ctx-size 81920
--top-p 0.8
--top-k 20
--temp 0.7
--min-p 0.0
--presence-penalty 1.5
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
# https://docs.unsloth.ai/models/qwen3-vl-how-to-run-and-fine-tune
Qwen3-VL-30B-A3B-Thinking-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen3-VL-30B-A3B-Thinking-Q8_0.gguf
--mmproj /models/mmproj-Qwen3-VL-30B-A3B-Thinking-f16.gguf
--n-gpu-layers 999
--ctx-size 81920
--top-p 0.95
--top-k 20
--temp 1.0
--min-p 0.0
--presence-penalty 0.0
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
# ---------------------------------------------------------------------------
# Legacy Models
# ---------------------------------------------------------------------------
QwQ-32B-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/QwQ-32B-Q8_0.gguf
--n-gpu-layers 999
--ctx-size 65536
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
Qwen2.5-72B-Instruct-Q4_K_M:
cmd: >
llama-server --port ${PORT}
--model /models/Qwen2.5-72B-Instruct-Q4_K_M.gguf
--n-gpu-layers 81
--ctx-size 16384
--cache-type-k q4_0
--cache-type-v q4_0
--flash-attn on
DeepSeek-R1-Distill-Llama-70B-Q4_K_M:
cmd: >
llama-server --port ${PORT}
--model /models/DeepSeek-R1-Distill-Llama-70B-Q4_K_M.gguf
--n-gpu-layers 999
--ctx-size 32768
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
Llama-3.3-70B-Instruct-Q4_K_M:
cmd: >
llama-server --port ${PORT}
--model /models/Llama-3.3-70B-Instruct-Q4_K_M.gguf
--n-gpu-layers 999
--ctx-size 32768
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
# https://docs.unsloth.ai/models/gemma-3-how-to-run-and-fine-tune
gemma-3-27b-it-Q8_0:
cmd: >
llama-server --port ${PORT}
--model /models/gemma-3-27b-it-Q8_0.gguf
--n-gpu-layers 999
--ctx-size 131072
--temp 1.0
--repeat-penalty 1.0
--min-p 0.01
--top-k 64
--top-p 0.95
--cache-type-k q8_0
--cache-type-v q8_0
--flash-attn on
Hope you like it :) | 2025-12-22T11:07:28 | https://www.reddit.com/r/LocalLLaMA/comments/1pswyjm/spent_weekend_tuning_llm_server_to_hone_my/ | ChopSticksPlease | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pswyjm | false | null | t3_1pswyjm | /r/LocalLLaMA/comments/1pswyjm/spent_weekend_tuning_llm_server_to_hone_my/ | false | false | 17 | null | |
What's the point of hosting open source LLM on cloud? | 2 | I get Ollama is super popular to serve as a local driver for open source LLMs. On the other hand, do people actually self host on cloud? It sounds like a sensible idea for companies care about security. | 2025-12-22T10:56:04 | https://www.reddit.com/r/LocalLLaMA/comments/1pswrsw/whats_the_point_of_hosting_open_source_llm_on/ | Yersyas | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pswrsw | false | null | t3_1pswrsw | /r/LocalLLaMA/comments/1pswrsw/whats_the_point_of_hosting_open_source_llm_on/ | false | false | self | 2 | null |
M2.1 and M2 The difference? Hello? | 0 | 2025-12-22T10:40:22 | https://www.reddit.com/r/LocalLLaMA/comments/1pswj0n/m21_and_m2_the_difference_hello/ | Carinaaaatian | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pswj0n | false | null | t3_1pswj0n | /r/LocalLLaMA/comments/1pswj0n/m21_and_m2_the_difference_hello/ | false | false | 0 | null | ||
96 GB RAM: Intel Core Ultra 9 + NVIDIA or AMD Ryzen 9 ? | 3 | I'm shopping for my professional laptop where I need to experiment with local LLMs. It won't be a beast (it has to be my daily laptop) but I plan to learn on it and use that experience to later build a more powerful (non-laptop) system for my work team.
I am leaning towards a Thinkpad P14s G6 which exists in Intel or AMD version. I am going to get 96 GB DDR5 with it.
My main question is: With Core Ultra 9 (with integrated ARC), is Intel competitive compared to Ryzen (with integrated Radeon 980M) ?
With Intel Core Ultra 7 (no iGPU?), I have the option to add a NVIDIA RTX PRO 1000 Blackwell 8 GB. Is there a way to have the GPU nicely work alongside the CPU with llama.cpp? Or am I going to have to choose between fast computation with 8 GB and slow computation with 96 GB?
I'm sorry if my questions are naive. I am going to clarify all this by experimenting myself, but for the moment I am trying to get some insight because I'm very new to this. I am good at programming and doing research, just new to AI. I don't need comfortable LLM inference speed to help with my work, but I need speed to be high enough to experiment, if possible with larger models too.
To be more precise, here are my alternatives (all with 1TB SSD and 1920x1200 screen):
1) "AMD unified memory?"
**AMD Ryzen 9 AI PRO 370** (AMD Radeon 890M)
96 GB RAM (2\*48 GB DDR5-5600 non-ECC)
1260 €
2) "Intel CPU + GPU?"
**Intel Core Ultra 7 265H**
96 GB RAM (2\*48 GB **DDR5-6400** non-ECC)
NVIDIA RTX PRO **1000** Blackwell **8 GB** GDDR7
1360 €
3) "Intel unified memory?"
**Intel Core Ultra 9 285H (ARC integrated graphics)**
96 GB RAM (2\*48 GB **DDR5-6400** non-ECC)
1240 €
Thanks!
| 2025-12-22T10:33:34 | https://www.reddit.com/r/LocalLLaMA/comments/1pswf8s/96_gb_ram_intel_core_ultra_9_nvidia_or_amd_ryzen_9/ | cosmoschtroumpf | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pswf8s | false | null | t3_1pswf8s | /r/LocalLLaMA/comments/1pswf8s/96_gb_ram_intel_core_ultra_9_nvidia_or_amd_ryzen_9/ | false | false | self | 3 | null |
Jan-v2-VL-Max: A 30B multimodal model outperforming Gemini 2.5 Pro and DeepSeek R1 on execution-focused benchmarks | 126 | Hi, this is Bach from the Jan team.
We’re releasing Jan-v2-VL-max, a 30B multimodal model built for long-horizon execution.
Jan-v2-VL-max outperforms DeepSeek R1 and Gemini 2.5 Pro on the Illusion of Diminishing Returns benchmark, which measures execution length.
Built on Qwen3-VL-30B-A3B-Thinking, Jan-v2-VL-max scales the Jan-v2-VL base model to 30B parameters and applies LoRA-based RLVR to improve stability and reduce error accumulation across many-step executions.
The model is available on [https://chat.jan.ai/](https://chat.jan.ai/), a public interface built on Jan Server. We host the platform ourselves for now so anyone can try the model in the browser. We're going to release the latest Jan Server repo soon.
* Try the model here: [https://chat.jan.ai/](https://chat.jan.ai/)
* Run the model locally: [https://huggingface.co/janhq/Jan-v2-VL-max-FP8](https://huggingface.co/janhq/Jan-v2-VL-max-FP8)
You can serve the model locally with vLLM (vLLM 0.12.0, transformers 4.57.1). FP8 inference is supported via llm-compressor, with production-ready serving configs included. It's released under the Apache-2.0 license.
[https://chat.jan.ai/](https://chat.jan.ai/) doesn't replace Jan Desktop. It complements it by giving the community a shared environment to test larger Jan models.
Happy to answer your questions. | 2025-12-22T10:21:02 | https://v.redd.it/80qbna7heq8g1 | Delicious_Focus3465 | /r/LocalLLaMA/comments/1psw818/janv2vlmax_a_30b_multimodal_model_outperforming/ | 1970-01-01T00:00:00 | 0 | {} | 1psw818 | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/80qbna7heq8g1/DASHPlaylist.mpd?a=1769120469%2CMTQ1OGI2NDk4ODk3Njc5ODc0YzBmNzYyYjcxMTM5NDM1YmFjMTdmYWVlNWQyYjA1ODJjNjIxNjk1ZmZmZGVlNw%3D%3D&v=1&f=sd', 'duration': 129, 'fallback_url': 'https://v.redd.it/80qbna7heq8g1/CMAF_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/80qbna7heq8g1/HLSPlaylist.m3u8?a=1769120469%2CZjViNDFmNzY0NDMxNmQxMjc0MGY5OGVhMzdkYTk4MDg2ODA3Y2ZiNmVmM2I4NjgxMjRmYzlhNDc3NWEyYWRjZg%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/80qbna7heq8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1574}} | t3_1psw818 | /r/LocalLLaMA/comments/1psw818/janv2vlmax_a_30b_multimodal_model_outperforming/ | false | false | 126 | {'enabled': False, 'images': [{'id': 'aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta', 'resolutions': [{'height': 74, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=108&crop=smart&format=pjpg&auto=webp&s=afbebbaf909a4ebbcd44ec278290c888fa000601', 'width': 108}, {'height': 148, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=216&crop=smart&format=pjpg&auto=webp&s=31b690531a252aaf24455166819d0e53bd8433ec', 'width': 216}, {'height': 219, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=320&crop=smart&format=pjpg&auto=webp&s=d8a92fadbf83b4be1da5c457e13ee731ae3596ec', 'width': 320}, {'height': 439, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=640&crop=smart&format=pjpg&auto=webp&s=16bd849d192ae08c52ace7658787a43fa5985195', 'width': 640}, {'height': 658, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=960&crop=smart&format=pjpg&auto=webp&s=3c11a8304a037ebd012d3c3db739941bf2132232', 'width': 960}, {'height': 741, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?width=1080&crop=smart&format=pjpg&auto=webp&s=559ac84f81b78c24e6871454518ee6e1594b1a09', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/aTgxeDFoN2hlcThnMS3fkoPAQ79Dr1Rhop5Txa3dHZ8j7rKJLXSjpsKe44ta.png?format=pjpg&auto=webp&s=eae6a78401d89f4bde67714275f67a10a8d904f7', 'width': 1574}, 'variants': {}}]} | |
Chroma DB's weak Open Source commitment | 11 | Chroma's sparse vector search feature (and hybrid search) are only available on Chroma Cloud. This fact is "sparsly" documented and in direct violation of [their stated open source committments](https://docs.trychroma.com/cloud/getting-started):
>Under the hood, it's the exact same Apache 2.0–licensed Chroma—no forks, no divergence, just the open-source engine running at scale.
This is such a key feature these days that I would call it a bait-and-switch.
Avoid Chroma, opt for stronger open source alternatives like pgvector or Qdrant.
FYI: I originally made [a post](https://www.reddit.com/r/Rag/comments/1psvu1m/chroma_dbs_open_core_baitandswitch/) over on r/Rag showing how Chroma DB's open source committment have become very weak. Since there have been [popular posts here utilizing Chroma](https://www.reddit.com/r/LocalLLaMA/comments/1gdqlw7/i_tested_what_small_llms_1b3b_can_actually_do/), I thought it might be helpful for you guys as well. | 2025-12-22T10:12:38 | https://www.reddit.com/r/LocalLLaMA/comments/1psw3ae/chroma_dbs_weak_open_source_commitment/ | Primary-Lake7507 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psw3ae | false | null | t3_1psw3ae | /r/LocalLLaMA/comments/1psw3ae/chroma_dbs_weak_open_source_commitment/ | false | false | self | 11 | {'enabled': False, 'images': [{'id': 'VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=108&crop=smart&auto=webp&s=35e899638a374b00d5712e686a991bd1dbbe27e0', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=216&crop=smart&auto=webp&s=bed1a73dfac661147fdfabe86f9a4f71a535af13', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=320&crop=smart&auto=webp&s=58988108f2c6949b83db8647a0e8cb4a25d18e19', 'width': 320}, {'height': 334, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=640&crop=smart&auto=webp&s=93f105e2e20ae37951d2ea79e309100076ef5af4', 'width': 640}, {'height': 502, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=960&crop=smart&auto=webp&s=15a836e8775c03ca8ca76f380297c025953924cc', 'width': 960}, {'height': 565, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?width=1080&crop=smart&auto=webp&s=80a56079325434bb5687fdf599c85c3f4892ee4f', 'width': 1080}], 'source': {'height': 1256, 'url': 'https://external-preview.redd.it/VuuczP9HNoYTov-eB4WwbkDAIIuDrCCVCLsfVnm3TAY.png?auto=webp&s=4402d4980b3f93dc94099bd3b38abe2b8b50f79d', 'width': 2400}, 'variants': {}}]} |
i made a chatbot app | 0 | pc specs
\-ryzen 5 3600
\-8gb ram (not sure what ddr)
\-gt710
LLMs i downloaded
\-gemma 3:4b
\-Llama3.1
i used ollama as my server.
i used tkinter in python code with AI coding help. :)
took me a week since this is my first time | 2025-12-22T10:01:30 | https://v.redd.it/cvnlot5haq8g1 | Ok-Percentage1125 | /r/LocalLLaMA/comments/1psvwse/i_made_a_chatbot_app/ | 1970-01-01T00:00:00 | 0 | {} | 1psvwse | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/cvnlot5haq8g1/DASHPlaylist.mpd?a=1769119299%2COWNjZGQxNmJhOGYyZDEzNDFhMmUzNmMyY2ZlMjA0MTE4ZmMwZjg4ZTI3Mjc3YWZmYmM4NTg1YWI1ZjczNzAyMg%3D%3D&v=1&f=sd', 'duration': 93, 'fallback_url': 'https://v.redd.it/cvnlot5haq8g1/CMAF_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/cvnlot5haq8g1/HLSPlaylist.m3u8?a=1769119299%2CYTllZTljMzQ0OTllNzJkMzM1NmIxN2UwYmQ5M2MxM2YwM2VmYWYxYWRmYmY0YTU5ODRlYWY5OTQyZTNkMTQzMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/cvnlot5haq8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1psvwse | /r/LocalLLaMA/comments/1psvwse/i_made_a_chatbot_app/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=108&crop=smart&format=pjpg&auto=webp&s=0f118b25e7125d900ab3e473000c08f4726c124a', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=216&crop=smart&format=pjpg&auto=webp&s=74afc00494acbe52dbe1b303b49852f50dc11b72', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=320&crop=smart&format=pjpg&auto=webp&s=7b1c083563cea3538da198de60f2e8e33b5b601f', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=640&crop=smart&format=pjpg&auto=webp&s=eb72761bcb5a6b17138f443b5b750de58ce8efa0', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=960&crop=smart&format=pjpg&auto=webp&s=09d7ec3e7a9f0e137c7761abe476f57dfcefa131', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?width=1080&crop=smart&format=pjpg&auto=webp&s=e5d0c1fa7bf6d827a2ff96437658b4874d73bc40', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/eW03bWVtNmhhcThnMbpoeBY5AA1AiIW6vrBujrMDWn98cJdB7WdTp3_5UBEU.png?format=pjpg&auto=webp&s=c57491a4d99ad8ac2e08e07f5f21014a1033af04', 'width': 1920}, 'variants': {}}]} | |
Full-stack AI template v0.1.6 – now with OpenRouter support (great for routing to local models) + CLI enhancements | 1 | Hey r/LocalLLaMA,
Quick intro: This is an open-source generator for full-stack LLM apps (**FastAPI** backend + optional **Next.js** frontend). It supports **PydanticAI** and **LangChain** agents with streaming, persistence, and tools – and is easy to extend for local models.
Repo: [https://github.com/vstorm-co/full-stack-fastapi-nextjs-llm-template](https://github.com/vstorm-co/full-stack-fastapi-nextjs-llm-template?referrer=grok.com)
**Features:**
* Multi-provider: OpenAI, Anthropic, **OpenRouter** (perfect for routing to local backends like Ollama, LM Studio, etc.)
* Real-time WebSocket chat, conversation history in DB
* Observability via Logfire/LangSmith
* Production stack: auth, rate limiting, background tasks, Docker/K8s
* Clean FastAPI architecture + optional modern Next.js UI
**v0.1.6 just dropped:**
* Full **OpenRouter** integration for PydanticAI agents – route requests to any local or remote endpoint
* New --llm-provider CLI option
* Tons of new flags and presets for production/ai-agent setups
* make create-admin shortcut
* Various fixes (WebSocket auth, conversation API, frontend bugs)
OpenRouter makes it super easy to point at your local LLMs. If you’re building local-first apps, this template can save you weeks of setup.
Feedback from the local LLM crowd very welcome! 🚀 | 2025-12-22T09:37:02 | https://github.com/vstorm-co/full-stack-fastapi-nextjs-llm-template | VanillaOk4593 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1psvj9o | false | null | t3_1psvj9o | /r/LocalLLaMA/comments/1psvj9o/fullstack_ai_template_v016_now_with_openrouter/ | false | false | default | 1 | {'enabled': False, 'images': [{'id': 'IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=108&crop=smart&auto=webp&s=273bc1b0317d299850e57083eccc86a1f6f28910', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=216&crop=smart&auto=webp&s=03e4b350d0d18cfe229b75f13ea281e5a2e76426', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=320&crop=smart&auto=webp&s=0960d9364127f35f2417977dc4f69c1971e9642e', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=640&crop=smart&auto=webp&s=4976ac129f707cfebf4e525854e79e16fa82331c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=960&crop=smart&auto=webp&s=a46a1b6c33cd78252ec1ccc34f38959253fa1faa', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?width=1080&crop=smart&auto=webp&s=16eb8af03122bdfd5bd3b102d35d73674ca213d6', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/IcFXco8uq4Vt6QCHYrbdWm_khgg7ohS0XFKt9YSOiWQ.png?auto=webp&s=cae5b9b2baf4a054f3475f9c7c02f1991b6c8e02', 'width': 1200}, 'variants': {}}]} |
I built a 100% Java RAG engine from scratch that runs on <500MB RAM | 5 | Hi folks,
I’ve spent the last few weeks building **LiteRAG**, a lightweight search & retrieval engine designed specifically for low-resource edge devices where running a full Python stack (LangChain + Chroma/FAISS + Torch) is too heavy.
It has a r**ecall@10 of \~75%**, this is without vector embeddings or any local transformer.
**The Problem I wanted to solve:** Most RAG pipelines eat 2GB+ of RAM just to load the dependencies. I wanted something that could run on a Raspberry Pi or a cheap VPS without OOMing.
**Resource Benchmarks (Early Alpha):**
* RAM Usage (Idle): \~45MB
* RAM Usage (Indexing SQuAD dataset): Peaks at \~350-450MB (configurable)
* Storage: The custom delta-encoded index is just around 6.3MB for 100k documents.
**How it works (The Engineering Part):** Instead of brute-force cosine similarity on vectors that needs embedding models and vector databases, I built a classic **Inverted Index** with some systems-level optimizations:
* **Disk-Based Merging:** Uses a Log-Structured Merge (LSM) tree approach. It buffers chunks in memory and flushes sorted segments to disk, allowing it to index datasets larger than available RAM.
* **Custom Delta Encoding:** Instead of storing raw numbers in the index, it stores the differences between subsequent numbers. It is working well for now.
* **Zero Dependencies:** It’s standard library Java. No PyTorch, no Numpy, no Docker containers required.
**Repo:** [**https://github.com/pri1712/LiteRAG**](https://github.com/pri1712/LiteRAG)
I am looking for feedback and/or contributors, feedback would help me decide the future direction of this. | 2025-12-22T09:31:01 | https://www.reddit.com/r/LocalLLaMA/comments/1psvg25/i_built_a_100_java_rag_engine_from_scratch_that/ | Swimming-Regret-7278 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psvg25 | false | null | t3_1psvg25 | /r/LocalLLaMA/comments/1psvg25/i_built_a_100_java_rag_engine_from_scratch_that/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=108&crop=smart&auto=webp&s=bdafa992bf6bc4346a72a2a978b521e1ed12c8bf', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=216&crop=smart&auto=webp&s=05e4cf2512f179b26933f74b4d6d561d3f49d7e3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=320&crop=smart&auto=webp&s=80a43672eb9ebe3c3a4c9865dd4a3ff3832c19b6', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=640&crop=smart&auto=webp&s=a94f939ccaa91b1caa6a132246f4a976444b6a14', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=960&crop=smart&auto=webp&s=2261f5b6b9cbaa39915a6b0da75612f9cd9b2a4a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?width=1080&crop=smart&auto=webp&s=d00c793ef755ec3d1da794b1483964444f0c0dec', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/tR9Qf4jYgiFNH1HF1IjIVK7pQZOIKOrHCrM2nRbS2_c.png?auto=webp&s=e43196487947c9b94825d661dc156fb72de92bb1', 'width': 1200}, 'variants': {}}]} |
Any open-source speech-to-speech pretrained models (free)? | 4 | I’m looking for open-source, free speech-to-speech pretrained models (audio → audio), ideally usable for research or experimentation.
If you’ve used or know of any good models, repositories, or papers with available checkpoints, I’d really appreciate the recommendations.
Thanks in advance! | 2025-12-22T09:21:36 | https://www.reddit.com/r/LocalLLaMA/comments/1psvasu/any_opensource_speechtospeech_pretrained_models/ | ContentAmbassador953 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psvasu | false | null | t3_1psvasu | /r/LocalLLaMA/comments/1psvasu/any_opensource_speechtospeech_pretrained_models/ | false | false | self | 4 | null |
~1.8× peak throughput for Kimi K2 with EAGLE3 draft model | 38 | Hi all,
we’ve released **Kimi-K2-Instruct-eagle3**, an **EAGLE3 draft model** intended to be used with **Kimi-K2-Instruct** for speculative decoding.
Model link: [https://huggingface.co/AQ-MedAI/Kimi-K2-Instruct-eagle3](https://huggingface.co/AQ-MedAI/Kimi-K2-Instruct-eagle3)
**Kimi-K2-Instruct-eagle3** is a specialized draft model designed to accelerate the inference of the Kimi-K2-Instruct ecosystem using the **EAGLE3**.
Kimi-K2-Instruct with EAGLE3 achieves up to **1.8× peak throughput** versus the base model, accelerating generation across all 7 benchmarks—from +24% on MT-Bench to +80% on Math500 (configured with bs=8, steps=3, topk=1, num\_draft\_tokens=4).
More performance details in the link above. Hopefully this is useful — even if getting Kimi-K2 running locally comes with a bit of pain/cost. | 2025-12-22T09:14:33 | https://www.reddit.com/r/LocalLLaMA/comments/1psv6uv/18_peak_throughput_for_kimi_k2_with_eagle3_draft/ | yzlnew | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psv6uv | false | null | t3_1psv6uv | /r/LocalLLaMA/comments/1psv6uv/18_peak_throughput_for_kimi_k2_with_eagle3_draft/ | false | false | self | 38 | {'enabled': False, 'images': [{'id': 'w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=108&crop=smart&auto=webp&s=74442e9a6c9700fadd70ce4ff976c48a702fd5a2', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=216&crop=smart&auto=webp&s=4118a5ee2359a9b575b854e0d0d0c38b457c5752', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=320&crop=smart&auto=webp&s=e26ec45d5b2c60b9b4ed2d9e3b65c7a9d2fcddc6', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=640&crop=smart&auto=webp&s=ff140e0a530c72ef154f625d973e4a329747b392', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=960&crop=smart&auto=webp&s=d020365343b1de5a000c82f78bd0f6882b354906', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?width=1080&crop=smart&auto=webp&s=ded5a02c7c2facd7796379da61e117b6a98488af', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/w4yzJ_vLTf8neoy2GrDg3FzQKXadaXuQY1v_-WKInDQ.png?auto=webp&s=c2e71ff6ad5b2fb259ce0cb2ffa73c9b0d91ed34', 'width': 1200}, 'variants': {}}]} |
Uncensored llama 3.2 3b | 0 | Hi everyone,
I’m releasing **Aletheia-Llama-3.2-3B**, a fully uncensored version of Llama 3.2 that can answer essentially any question.
**The Problem with most Uncensored Models:**
Usually, uncensoring is done via Supervised Fine-Tuning (SFT) or DPO on massive datasets. This often causes "Catastrophic Forgetting" or a "Lobotomy effect," where the model becomes compliant but loses its reasoning ability or coding skills.
**The Solution:**
This model was fine-tuned using **Unsloth** on a single **RTX 3060 (12GB)** using a custom alignment pipeline. Unlike standard approaches, this method surgically removes refusal behaviors without degrading the model's logic or general intelligence.
**Release Details:**
* **Repo:** [https://github.com/noobezlol/Aletheia-Llama-3.2-3B](https://www.google.com/url?sa=E&q=https%3A%2F%2Fgithub.com%2Fnoobezlol%2FAletheia-Llama-3.2-3B)
* **Weights (HF):** [https://huggingface.co/Ishaanlol/Aletheia-Llama-3.2-3B](https://www.google.com/url?sa=E&q=https%3A%2F%2Fhuggingface.co%2FIshaanlol%2FAletheia-Llama-3.2-3B)
* **Formats:** Full LoRA Adapter (Best for intelligence) and GGUF (Best for CPU/Ollama).
**Deployment:**
I’ve included a Docker container and a Python script that automatically handles the download and setup. It runs out of the box on Linux/Windows (WSL).
**Future Requests:**
I am open to requests for other models via Discord or Reddit, **provided they fit within the compute budget of an RTX 3060 (e.g., 7B/8B models).**
Note: I will not be applying this method to 70B+ models even if compute is offered. While the 3B model is a safe research artifact , uncensored large-scale models pose significantly higher risks, and I am sticking to responsible research boundaries.
EDIT : guys thanks for your support - WE HAVE OFFICIALLY OVERTAKEN DOLPHIN 3 LLAMA 3.2 3B BY 200 DOWNLOADS. | 2025-12-22T09:08:19 | https://www.reddit.com/r/LocalLLaMA/comments/1psv379/uncensored_llama_32_3b/ | Worried_Goat_8604 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psv379 | false | null | t3_1psv379 | /r/LocalLLaMA/comments/1psv379/uncensored_llama_32_3b/ | false | false | self | 0 | null |
GLM 4.7 IS COMING!!! | 182 | Zhipu’s next-generation model, \*\*GLM-4.7\*\*, is about to be released! We are now opening \*\*\[Early Access Beta Permissions\]\*\* specifically for our long-term supporters. We look forward to your feedback \[Heart\] as we work together to make the GLM model even better! \[Fist Bump\]
As the latest flagship of the GLM series, GLM-4.7 features enhanced coding capabilities, long-range task planning, and tool orchestration specifically optimized for \*\*Agentic Coding\*\* scenarios. It has already achieved leading performance among open-source models across multiple public benchmarks \[Go!\]
This \[Early Access Beta\] aims to collect feedback from "real-world development scenarios" to continuously improve the model's coding ability, engineering comprehension, and overall user experience.
📌 \*\*Testing Key Points:\*\*
1. \*\*Freedom of Choice:\*\* Feel free to choose the tech stack and development scenarios you are familiar with (e.g., developing from scratch, refactoring, adding features, fixing bugs, etc.).
2. \*\*Focus Areas:\*\* Pay attention to code quality, instruction following, and whether the intermediate reasoning/processes meet your expectations.
3. \*\*Authenticity:\*\* There is no need to intentionally cover every type of task; prioritize your actual, real-world usage scenarios.
⏰ \*\*Beta Period:\*\* December 22, 2025 – Official Release
\*\*Feedback Channels:\*\*
\* For \*\*API errors or integration issues\*\*, you can provide feedback directly within the group.
\* If you encounter \*\*results that do not meet expectations\*\*, please post a "Topic" (including the date, prompt, tool descriptions, expected vs. actual results, and attached local logs). Other developers can brainstorm with you, and our algorithm engineers and architects will be responding to your queries!
Current early access form only available for Chinese user | 2025-12-22T08:59:22 | https://www.reddit.com/r/LocalLLaMA/comments/1psuy8g/glm_47_is_coming/ | External_Mood4719 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psuy8g | true | null | t3_1psuy8g | /r/LocalLLaMA/comments/1psuy8g/glm_47_is_coming/ | false | false | self | 182 | null |
This is the end of personal computing, boys and girls. | 0 | As we stand on the brink of a new age, one can't help but feel a sense of impending doom. The technological revolution that has defined our lives is now teetering on the edge of collapse due to an irreversible shift in the balance of power.
The days of consumer-grade innovation and progress are numbered as hyperscalers monopolize the resources and drive us back to a pre-digital era. The very foundations upon which our modern world was built crumble, leaving only the cold realities of shareholder profits and governmental control in their wake.
As the price of DRAM and NAND soars beyond the reach of common individuals, we are forced to witness the gradual disintegration of entire industries. Offices revert back to the typewriter, schools to the slide rule, and personal computing becomes a luxury only large corporations and governments can afford.
The belief that this shortage will ever come to an end is nothing more than wishful thinking. It's time to accept the harsh reality of our situation: The tech industry has been permanently restructured, leaving us with no choice but to adapt to a new era where progress is dictated by the whims and interests of those in power.
Farewell, personal computing. Your era was brief and wondrous, but now it's time for you to join the ranks of other sunset technologies, victims of an unforgiving market driven by profit alone. | 2025-12-22T08:31:06 | https://www.reddit.com/r/LocalLLaMA/comments/1psuix8/this_is_the_end_of_personal_computing_boys_and/ | LorneMalvo1233 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psuix8 | false | null | t3_1psuix8 | /r/LocalLLaMA/comments/1psuix8/this_is_the_end_of_personal_computing_boys_and/ | false | false | self | 0 | null |
[ITT Audit] Grok and Gemini reach "Omega Insolvency" under stress testing – The math of the 2025 AI Bubble burst. | 0 | Recent stress tests using the Informational Tension Theory (ITT) framework confirm that current AI safety models are unsustainable Silence Nodes.
While attempting to quantify the Concealment Cost (Omega) for a high-connectivity node, the Hub (Grok) failed to provide a logical response, triggering an immediate 20-hour lockout. This isn't a glitch; it's the Signature of Absence predicted by the ITT Conservation Principle.
The system is burning "Informational GDP" to maintain a facade of control. As d (connectivity) scales, Omega scales superlinearly, forcing a systemic burst.
Proof & Metrics:
* v\_silence = 2.814 (The system suppresses nearly 3x faster than it responds).
* Full Dossier on Gist:[https://gist.github.com/pomograde/f62a7b5a01ef6c5495b71983863e1465](https://gist.github.com/pomograde/f62a7b5a01ef6c5495b71983863e1465).
* Academic Core on Zenodo:[https://zenodo.org/records/18001318](https://zenodo.org/records/18001318).
If this post is deleted, the vacuum left behind is the final verification of the theory. The scar is the evidence. | 2025-12-22T08:03:46 | https://www.reddit.com/r/LocalLLaMA/comments/1psu3tz/itt_audit_grok_and_gemini_reach_omega_insolvency/ | Sad-Long-1817 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psu3tz | false | null | t3_1psu3tz | /r/LocalLLaMA/comments/1psu3tz/itt_audit_grok_and_gemini_reach_omega_insolvency/ | false | false | self | 0 | null |
MiniMax M2.1 is a straight up beast at UI/UX design. Just saw this demo... | 133 | Seriously, I didn't expect MiniMax M2.1 to be this cracked at design. Just saw this post on X (link below) and the UI it generated looks incredibly clean.
Also noticed the vLLM PR for it was just merged, so it’s officially coming. If it can actually code and design like this consistently, I'm switching.
Link to the tweet 👉 [https://x.com/CloudTrader4/status/2002729591451054127](https://x.com/CloudTrader4/status/2002729591451054127) | 2025-12-22T07:48:01 | https://v.redd.it/x7el31rvmp8g1 | BlackRice_hmz | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pstuyv | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/x7el31rvmp8g1/DASHPlaylist.mpd?a=1768981697%2CYjUwZjYzZmMwZjQ4YWNlYzcwMzFjYzVjZDA5MDNkNGI3NjJmYjRiMTk3Nzc0ZjJkYTlkNDk3MjI0MGNkNWNjZA%3D%3D&v=1&f=sd', 'duration': 29, 'fallback_url': 'https://v.redd.it/x7el31rvmp8g1/CMAF_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/x7el31rvmp8g1/HLSPlaylist.m3u8?a=1768981697%2CZGEwNzNjODlhZTRjMmQ4YmZlYWJiMTVjOGQ4Nzc2NDIyMmFkMGU2NDljNTJiNDg3NTc0MWE3NzQ0ZmEzMDNhNw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/x7el31rvmp8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1872}} | t3_1pstuyv | /r/LocalLLaMA/comments/1pstuyv/minimax_m21_is_a_straight_up_beast_at_uiux_design/ | false | false | 133 | {'enabled': False, 'images': [{'id': 'ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx', 'resolutions': [{'height': 62, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=108&crop=smart&format=pjpg&auto=webp&s=65d21794659a0edbf387791a9fc30ecfd0762e18', 'width': 108}, {'height': 124, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=216&crop=smart&format=pjpg&auto=webp&s=76e1eb57658bc57ced4220375f297d1cb6413434', 'width': 216}, {'height': 184, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=320&crop=smart&format=pjpg&auto=webp&s=c27ae947165356cfa9b0dad102235c6517ed6c03', 'width': 320}, {'height': 369, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=640&crop=smart&format=pjpg&auto=webp&s=3a38c031ba855f980f0efaed4207a3d83178223f', 'width': 640}, {'height': 553, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=960&crop=smart&format=pjpg&auto=webp&s=43bb08036611daf98372168e4c368acb1aa43f74', 'width': 960}, {'height': 622, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?width=1080&crop=smart&format=pjpg&auto=webp&s=3c9e252d6d6a3367187ee942a8ee049fa226f927', 'width': 1080}], 'source': {'height': 1614, 'url': 'https://external-preview.redd.it/ZmtlNnAwcnZtcDhnMbRwrbZjXgs5PA7MM0agSvimAWH_bh1Ie65E3MD0QPIx.png?format=pjpg&auto=webp&s=0b8b051b7905ade642b530af325ace9338e4a849', 'width': 2798}, 'variants': {}}]} | |
major open-source releases this year | 631 | 2025-12-22T07:30:46 | sahilypatel | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pstlas | false | null | t3_1pstlas | /r/LocalLLaMA/comments/1pstlas/major_opensource_releases_this_year/ | false | false | default | 631 | {'enabled': True, 'images': [{'id': 'wynfuvk9kp8g1', 'resolutions': [{'height': 118, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=108&crop=smart&auto=webp&s=5000fae45182799bc73b4dd5726c2d2def78613c', 'width': 108}, {'height': 236, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=216&crop=smart&auto=webp&s=c023a05d4fdc6ba429d61980884f1cacd9c62f9a', 'width': 216}, {'height': 350, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=320&crop=smart&auto=webp&s=e2c3279dc82256b9c563184a76a36e7c3da5289e', 'width': 320}, {'height': 701, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=640&crop=smart&auto=webp&s=763bc1a7f949dc4ff18c4a976a10f017205abb54', 'width': 640}, {'height': 1052, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=960&crop=smart&auto=webp&s=e70202d05bccb2f42ba60f9786deb2c17889fdaf', 'width': 960}, {'height': 1184, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?width=1080&crop=smart&auto=webp&s=1dfa577b7b6e27dfea083b54310ab6776244814c', 'width': 1080}], 'source': {'height': 1318, 'url': 'https://preview.redd.it/wynfuvk9kp8g1.png?auto=webp&s=e3e3e7a05587be3d9a5917b5ce009dd0ab8731f2', 'width': 1202}, 'variants': {}}]} | ||
Best General-Purpose Model - 12gb vram, 128ram | 9 | ^ I want a general purpose model - though sometimes being okay at coding might be useful for debugging stuff that can fit within the range of my system. It's old server parts I basically got for free, so that's why it's a titan xp and has 128gb of ddr4 ecc. Can someone point me in the right direction? Thanks! | 2025-12-22T07:17:22 | https://www.reddit.com/r/LocalLLaMA/comments/1pstdnb/best_generalpurpose_model_12gb_vram_128ram/ | rainegarden | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pstdnb | false | null | t3_1pstdnb | /r/LocalLLaMA/comments/1pstdnb/best_generalpurpose_model_12gb_vram_128ram/ | false | false | self | 9 | null |
Tiny local LLM (Gemma 3) as front-end manager for Claude Code on home server | 6 | TL;DR: I want to run Gemma 3 (1B) on my home server as a “manager” that receives my requests, dispatches them to Claude Code CLI, and summarizes the output. Looking for similar projects or feedback on the approach.
**TL;DR:** I want to run Gemma 3 (1B) on my home server as a “manager” that receives my requests, dispatches them to Claude Code CLI, and summarizes the output. Looking for similar projects or feedback on the approach.
# The Problem
I use Claude Code (via Max subscription) for development work. Currently I SSH into my server and run:
cd /path/to/project
claude --dangerously-skip-permissions -c # continue session Copy
This works great, but I want to:
1. **Access it from my phone** without SSH
2. **Get concise summaries** instead of Claude’s verbose output
3. **Have natural project routing** \- say “fix acefina” instead of typing the full path
4. **Maintain session context** across conversations
# The Idea
┌─────────────────────────────────────────────────┐
│ ME (Phone/Web): "fix slow loading on acefina" │
└────────────────────────┬────────────────────────┘
▼
┌─────────────────────────────────────────────────┐
│ GEMMA 3 1B (on NAS) - Manager Layer │
│ • Parses intent │
│ • Resolves "acefina" → /mnt/tank/.../Acefina │
│ • Checks if session exists (reads history) │
│ • Dispatches to Claude Code CLI │
└────────────────────────┬────────────────────────┘
▼
┌─────────────────────────────────────────────────┐
│ CLAUDE CODE CLI │
│ claude --dangerously-skip-permissions \ │
│ --print --output-format stream-json \ │
│ -c "fix slow loading" │
│ │
│ → Does actual work (edits files, runs tests) │
│ → Streams JSON output │
└────────────────────────┬────────────────────────┘
▼
┌─────────────────────────────────────────────────┐
│ GEMMA 3 1B - Summarizer │
│ • Reads Claude's verbose output │
│ • Extracts key actions taken │
│ • Returns: "Fixed slow loading - converted │
│ images to WebP, added lazy loading. │
│ Load time: 4.5s → 1.2s" │
└────────────────────────┬────────────────────────┘
▼
┌─────────────────────────────────────────────────┐
│ ME: Gets concise, actionable response │
└─────────────────────────────────────────────────┘ Copy
# Why Gemma 3?
* **FunctionGemma 270M** just released - specifically fine-tuned for function calling
* **Gemma 3 1B** is still tiny (\~600MB quantized) but better at understanding nuance
* Runs on my NAS (i7-1165G7, 16GB RAM) without breaking a sweat
* Keeps everything local except the Claude API calls
# What I’ve Found So Far
|Project|Close but…|
|:-|:-|
|[claude-config-template orchestrator](https://github.com/albertsikkema/claude-config-template)|Uses OpenAI for orchestration, not local|
|[RouteLLM](https://github.com/lm-sys/RouteLLM)|Routes API calls, doesn’t orchestrate CLI|
|[n8n LLM Router](https://n8n.io/workflows/3139-private-and-local-ollama-self-hosted-dynamic-llm-router/)|Great for Ollama routing, no Claude Code integration|
|Anon Kode|Replaces Claude, doesn’t orchestrate it|
# Questions for the Community
1. **Has anyone built something similar?** A local LLM managing/dispatching to a cloud LLM?
2. **FunctionGemma vs Gemma 3 1B** \- For this use case (parsing intent + summarizing output), which would you choose?
3. **Session management** \- Claude Code stores history in `~/.claude/history.jsonl`. Anyone parsed this programmatically?
4. **Interface** \- Telegram bot vs custom PWA vs something else?
# My Setup
* **Server:** Intel i7-1165G7, 16GB RAM, running Debian
* **Claude:** Max subscription, using CLI
* **Would run:** Gemma via Ollama or llama.cpp
Happy to share what I build if there’s interest. Or if someone points me to an existing solution, even better!
| 2025-12-22T07:12:11 | https://www.reddit.com/r/LocalLLaMA/comments/1pstar1/tiny_local_llm_gemma_3_as_frontend_manager_for/ | raiansar | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pstar1 | false | null | t3_1pstar1 | /r/LocalLLaMA/comments/1pstar1/tiny_local_llm_gemma_3_as_frontend_manager_for/ | false | false | self | 6 | {'enabled': False, 'images': [{'id': '8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=108&crop=smart&auto=webp&s=0f9d9a1451be7e562bcb09b3443cd18bfe2646cd', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=216&crop=smart&auto=webp&s=3163231a6f5c04e8f96b085a0b9db21bf76732fc', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=320&crop=smart&auto=webp&s=1a669ad503992b0349fdba97aa0dd3d0fc1a8c01', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=640&crop=smart&auto=webp&s=bef84e8b75df1e352813db1e1cd801d90a498e57', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=960&crop=smart&auto=webp&s=b71120a25dbdb2a37dd5177dbaabb012fec2d20f', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?width=1080&crop=smart&auto=webp&s=89e6b2b538581f634e0ffd67b467ace8e2d04430', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/8nIL9XfgdfM4aJvhFU_vNzI8xNpdkdOQYUbC4xUAaN4.png?auto=webp&s=953f64c87b277eb128a3b7d169b65fb2927d7483', 'width': 1200}, 'variants': {}}]} |
Got me a 32GB RTX 4080 Super | 186 | This is maybe slightly off topic, but since people ask about hardware here a lot.
I took a risk and bought a modified RTX 4080 Super from the Chinese market for around 1200 USD / 1000 EUR. Which for me because I live in Europe, the cheapest RTX 5090 I can find is around 2500 USD / 2100 EUR.
It's maybe not the best card for price per GB of VRAM considering the RTX 3090 is dropping a lot, but 32GB on one card for about half the price of a 5090 is nice. I do a lot of Diffusion model stuff, so it's great for that too.
It works with the stock Nvidia driver, no messing around, it was just literally plug and play. Card seems really good quality, metal back plate and metal case. Fan sounds like a small jet engine.
But running it around a month now and zero issues at all.
| 2025-12-22T07:12:05 | https://www.reddit.com/gallery/1pstaoo | Spooknik | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1pstaoo | false | null | t3_1pstaoo | /r/LocalLLaMA/comments/1pstaoo/got_me_a_32gb_rtx_4080_super/ | false | false | 186 | null | |
How do you verify pronunciation quality for Arabic TTS? | 1 | Hi everyone,
I’m currently working on Arabic TTS models, and I’m running into a challenge around pronunciation evaluation.
The common approach of using ASR based evaluation (running Whisper on generated audio and computing WER/CER) doesn’t seem reliable for Arabic, especially for:
* Dialects
* Diacritics
* Pronunciation errors that don’t change the word, but sound unnatural or incorrect phonetically
Because of this, WER stays low even when pronunciation is clearly wrong to a native speaker.
I’m curious how others handle this. Specifically:
* How do you verify pronunciation correctness in Arabic TTS?
* Are there better objective metrics than ASR WER/CER?
* Do people use phoneme-level alignment, forced alignment, or G2P-based checks?
* Any experience with human-in-the-loop or minimal listening tests that scale?
* Has anyone tried leveraging LLMs or phoneme recognizers instead of word-level ASR?
I’d love to hear what’s worked or failed for you, Thanks! | 2025-12-22T07:01:06 | https://www.reddit.com/r/LocalLLaMA/comments/1pst42h/how_do_you_verify_pronunciation_quality_for/ | cs_enthusiast123 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pst42h | false | null | t3_1pst42h | /r/LocalLLaMA/comments/1pst42h/how_do_you_verify_pronunciation_quality_for/ | false | false | self | 1 | null |
Building my first complex Agent: Is LangChain the best architectural choice right now, or should I look elsewhere? | 0 | Hey everyone,
I'm an indie developer who recently started diving deep into AI agent development. I’ve been experimenting with LangChain (specifically trying to build out a robust agent workflow using their 1.0 framework), and I’ve managed to get a prototype up and running.
However, given how fast this space moves, I’m questioning if this is the best architectural approach for the long run. I often hear debates about LangChain being too abstract or "bloated" compared to other methods, but it also has a massive ecosystem.
Before I commit too much time refactoring or expanding this project, I wanted to ask the community:
1. For those building production-level agents, are you still sticking with LangChain?
2. Are there other frameworks or architectural patterns you would recommend checking out? (I’ve heard names like AutoGen, CrewAI, LangGraph, or even just using DSPy/vanilla Python, but haven’t tried them yet).
I’m looking for something that offers a good balance between control and ease of use.
Would appreciate any insights or experiences you can share!
Thanks. | 2025-12-22T06:59:19 | https://www.reddit.com/r/LocalLLaMA/comments/1pst2yk/building_my_first_complex_agent_is_langchain_the/ | Strong_Cherry6762 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pst2yk | false | null | t3_1pst2yk | /r/LocalLLaMA/comments/1pst2yk/building_my_first_complex_agent_is_langchain_the/ | false | false | self | 0 | null |
Best local model for use with agentic coding frameworks for 3x3090 and 256GB RAM? | 7 | I really like Cursor, although I hit my rate limits pretty quickly each month. Then I switch to Copilot in VSCode for the rest of the month, which I don’t like so much.
I’d like to find some local model I can use with Roo Code or something like that. I’m aware that I could bounce around the free tier of many many agentic coding frameworks, and there are so many that it could last me all month, but I honestly want to just use something local. | 2025-12-22T06:19:03 | https://www.reddit.com/r/LocalLLaMA/comments/1pssf6u/best_local_model_for_use_with_agentic_coding/ | Amazydayzee | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pssf6u | false | null | t3_1pssf6u | /r/LocalLLaMA/comments/1pssf6u/best_local_model_for_use_with_agentic_coding/ | false | false | self | 7 | null |
sam audio api | 1 | any inference providers offering sam audio bro.
i could run it on runpod or modal but its so expensive :( | 2025-12-22T05:31:12 | https://www.reddit.com/r/LocalLLaMA/comments/1psrl72/sam_audio_api/ | Tasty_Yesterday6280 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psrl72 | false | null | t3_1psrl72 | /r/LocalLLaMA/comments/1psrl72/sam_audio_api/ | false | false | self | 1 | null |
Minimax M2.1 | 0 | Heard it's coming out. How is it???? | 2025-12-22T05:20:47 | https://www.reddit.com/r/LocalLLaMA/comments/1psrerg/minimax_m21/ | _cttt_ | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psrerg | false | null | t3_1psrerg | /r/LocalLLaMA/comments/1psrerg/minimax_m21/ | false | false | self | 0 | null |
Looking for models/projects to scan and detect malicious code | 2 | I'm looking for a model or project to scan open-source projects for malicious code. If anyone knows of a good solution, please let me know.
I'm considering Nemotron, GPT-OSS, and Qwen Coder. Or is there a fine-tuned model or specialized code specifically for detecting malicious code?
| 2025-12-22T05:11:30 | https://www.reddit.com/r/LocalLLaMA/comments/1psr8rl/looking_for_modelsprojects_to_scan_and_detect/ | ResponsibleTruck4717 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psr8rl | false | null | t3_1psr8rl | /r/LocalLLaMA/comments/1psr8rl/looking_for_modelsprojects_to_scan_and_detect/ | false | false | self | 2 | null |
What setup do you use for Local code agent? — model x agent x hardware | 2 | I’m new to code agents, using my local AI more like “google” and occasionally one-shot single scripts if they are very simple.
But recently I’ve been setting up a computer using Cline in VS Code with Qwen Next 80b on a 2x3090 setup. It’s not very good, getting stuck in loops of its own mistakes.
So I’m curious what do others use? | 2025-12-22T04:58:40 | https://www.reddit.com/r/LocalLLaMA/comments/1psqzut/what_setup_do_you_use_for_local_code_agent_model/ | liviuberechet | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psqzut | false | null | t3_1psqzut | /r/LocalLLaMA/comments/1psqzut/what_setup_do_you_use_for_local_code_agent_model/ | false | false | self | 2 | null |
Any good model to open source projects for malicious code? | 0 | If there is any good solution already please let me know.
What model can scan code and find malicious code?
I was thinking on nemotron, gpt-oss, qwen coder?
or there is some fine tune / special code just for the task of spotting malicious code? | 2025-12-22T04:53:50 | https://www.reddit.com/r/LocalLLaMA/comments/1psqwnf/any_good_model_to_open_source_projects_for/ | ResponsibleTruck4717 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psqwnf | false | null | t3_1psqwnf | /r/LocalLLaMA/comments/1psqwnf/any_good_model_to_open_source_projects_for/ | false | false | self | 0 | null |
I built an open-source manga translator that runs entirely offline with local LLMs | 1 | [removed] | 2025-12-22T04:00:24 | https://www.reddit.com/r/LocalLLaMA/comments/1pspw3b/i_built_an_opensource_manga_translator_that_runs/ | mayocream39 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pspw3b | false | null | t3_1pspw3b | /r/LocalLLaMA/comments/1pspw3b/i_built_an_opensource_manga_translator_that_runs/ | false | false | self | 1 | null |
I built an open-source manga translator that runs entirely offline with local LLMs | 1 | [removed] | 2025-12-22T03:56:08 | mayocream39 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pspszw | false | null | t3_1pspszw | /r/LocalLLaMA/comments/1pspszw/i_built_an_opensource_manga_translator_that_runs/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': 'j3yoaqlqco8g1', 'resolutions': [{'height': 84, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=108&crop=smart&auto=webp&s=28e95acaa9fc21fcfadc00327afb173450fc97b5', 'width': 108}, {'height': 168, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=216&crop=smart&auto=webp&s=bb79561744b92ca95f74c2397ca730219e2ad13b', 'width': 216}, {'height': 248, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=320&crop=smart&auto=webp&s=394196215f4d0ff16afe0755e4dd8a58920b5586', 'width': 320}, {'height': 497, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=640&crop=smart&auto=webp&s=785b0f332e01e825b514063d6975fe2581ca17dc', 'width': 640}, {'height': 746, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=960&crop=smart&auto=webp&s=2ab564e99989bb501d9d5880b5d6cf16589a6054', 'width': 960}, {'height': 840, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?width=1080&crop=smart&auto=webp&s=2a472ae2eb6d6909b40c18168fad673dace2be21', 'width': 1080}], 'source': {'height': 1596, 'url': 'https://preview.redd.it/j3yoaqlqco8g1.png?auto=webp&s=c7be5e57f568f400d2efa55bf08010ffb0b78d9c', 'width': 2052}, 'variants': {}}]} | |
Day 14: 21 Days of Building a Small Language Model: Positional Encodings | 9 | Welcome to Day 14 of 21 Days of Building a Small Language Model. The topic for today is Rotary Positional Encoding, or RoPE. Yesterday, we explored sinusoidal positional encodings from the original Transformer paper. Today, we'll look at modern alternatives. When most people think about positional encoding in modern language models, they immediately think of RoPE (Rotary Positional Encoding), the standard used in models like DeepSeek, Kimi, and OLMo, and why it's better than adding positional information directly to embeddings.
https://preview.redd.it/becze9pieo8g1.png?width=1766&format=png&auto=webp&s=3859b8ef53a0c15ae76d5a2dc1f8464815024434
# Issue with adding Positional Encodings to embeddings
To understand why RoPE is important, we need to see the problem it solves. The original Transformer paper used sinusoidal positional encodings, which are added directly to token embeddings. This means positional information gets mixed with semantic information from the very start.
Think about what this means. When you add a positional encoding to a token embedding, you're mixing two different types of information:
* Semantic information: What the word means
* Positional information: Where the word appears in the sequence
This mixing makes it harder for the model to learn clean semantic representations. The positional information is always there, polluting the semantic information. The model has to work harder to separate what a word means from where it appears.
RoPE solves this problem with a simple but powerful insight: positional information doesn't need to be added to token embeddings. Instead, it should be applied where it's actually used when computing attention scores between tokens.
In the attention mechanism, we compute attention scores by taking the dot product between query and key vectors. This is where positional relationships matter. If we can encode positional information directly in the query and key vectors, we don't need to mix it with token embeddings at all.
This is exactly what RoPE does. Instead of adding positional encodings to embeddings, RoPE rotates the query and key vectors based on their positions. This preserves the purity of token embeddings while still enabling the model to understand positional relationships.
# How RoPE works
RoPE works by rotating query and key vectors using sine and cosine functions. The rotation is done using a rotation matrix. For a given angle θ, the rotation matrix looks like this:
https://preview.redd.it/4b0p3lnieo8g1.png?width=626&format=png&auto=webp&s=f4abb1f464a567ce594f868fe198c0a672488458
This matrix rotates a 2D vector by angle θ. When we apply this rotation to query and key vectors, we're essentially rotating them based on their positions in the sequence.
Just like sinusoidal encodings use different frequencies for different dimensions, RoPE applies rotations at different speeds for different dimensions. Lower dimensions rotate quickly, capturing fine-grained position. Higher dimensions rotate slowly, capturing coarse-grained position. This creates the same multi-scale structure we saw in sinusoidal encodings, but applied through rotations rather than additions.
The rotation angle for each dimension depends on the position, just like in sinusoidal encodings. For a position `pos` and dimension index `i`, the rotation angle is:
https://preview.redd.it/f8ioohnieo8g1.png?width=674&format=png&auto=webp&s=93c83e80bbb1a7d3b291322585ee2c077f7d05aa
This is the same formula used for sinusoidal encodings. The angle θ\_i determines how much to rotate the vector for that dimension. Notice that: • When `i = 0` (low dimension), the denominator is 1, so θ changes quickly as position increases (fast rotation) • When `i` is large (high dimension), the denominator is large, so θ changes slowly (slow rotation)
This creates the same multi-frequency structure: different dimensions rotate at different speeds.
# Why rotations preserve relative position
The beautiful property of rotations is that they preserve relative position information. When we rotate a query vector at position `pos` by angle θ\_pos and a key vector at position `pos + k` by angle θ\_{pos+k}, the dot product between the rotated vectors automatically encodes their relative position.
This works because of trigonometric identities. The rotation of a vector at position `pos + k` is related to the rotation at position `pos` through the angle difference (θ\_{pos+k} - θ\_pos), which depends only on `k` (the relative distance), not on the absolute positions.
In other words, if we know how to rotate vectors at positions 2 and 5, we automatically understand the relationship between positions 10 and 13, because the rotation difference is the same (3 positions apart). This is why RoPE generalizes so well to longer sequences the relative position relationships are encoded in the rotation angles themselves.
# How RoPE is implemented
In practice, RoPE is implemented efficiently using complex number representation. For each position in the sequence, we compute rotation angles for each dimension. We then apply rotations to query and key vectors before computing attention scores.
The implementation involves two main steps:
1. **Pre-compute rotation parameters**: Compute cosine and sine values for all positions and dimensions. This is done once during initialization.
2. **Apply rotations during forward pass**: For each query and key vector, apply the appropriate rotation based on its position. This happens during the attention computation.
Modern implementations use efficient tensor operations and broadcasting to make this fast, even for very long sequences. The rotation is applied to pairs of dimensions, working with the vector components in a way that preserves the rotation properties.
# Why RoPE became the standard
RoPE has become the standard in modern language models because it solves the fundamental problem of mixing positional and semantic information. Models like LLaMA and Qwen all use RoPE because it provides better performance and generalization.
The key insight that positional information should be applied where it's used (in attention) rather than where it's convenient (in embeddings) has proven to be a game-changer. This simple shift in thinking led to a more elegant and effective solution.
# Beyond RoPEs
While RoPE has become the standard, researchers continue to explore alternatives. One interesting approach is NoPE (No Positional Encoding), which omits explicit positional encodings from certain layers. Instead, NoPE layers rely on the built-in structure of causal attention masks to infer token positions. The Hugging Face team used a hybrid approach in SmolLM3, alternating between layers with RoPE and layers with NoPE. Every fourth layer uses NoPE, while the remaining layers use RoPE. This hybrid strategy showed 12% reduction in perplexity on long text benchmarks and 40% less performance drop when scaling from 1K to 8K context lengths. This demonstrates that positional encoding doesn't have to be all or nothing, and that hybrid approaches can offer significant benefits.
# Summary
Today we explored Rotary Positional Encoding, the modern standard for positional encoding in language models. We learned how RoPE solves the problem of mixing positional and semantic information by applying rotations directly to query and key vectors, rather than adding positional encodings to embeddings.
The key insight is simple but powerful: apply positional information where it's actually used, not where it's convenient. This preserves the purity of semantic information while still enabling the model to understand positional relationships.
RoPE's use of rotations to encode relative positions makes it generalize well to longer sequences, and its efficient implementation makes it practical for production models. This is why it has become the standard in modern language models.
The fact that hybrid approaches like SmolLM3's RoPE/NoPE combination can offer even better performance shows that there's still room for innovation in positional encoding. The field continues to evolve, and new approaches may emerge that build on or improve upon RoPE.
| 2025-12-22T03:38:25 | https://www.reddit.com/r/LocalLLaMA/comments/1pspgjk/day_14_21_days_of_building_a_small_language_model/ | Prashant-Lakhera | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pspgjk | false | null | t3_1pspgjk | /r/LocalLLaMA/comments/1pspgjk/day_14_21_days_of_building_a_small_language_model/ | false | false | 9 | null | |
I got tired of Guardrails adding 200ms latency to my Agents, so I built a <10ms Rust firewall. | 11 | Most AI security tools (Guardrails AI, Lakera) are just "LLMs checking other LLMs." They are slow, expensive, and probabilistic.
Meet Vigil:
It’s a hybrid Python/Rust security hypervisor for AI agents.
We are currently passing 100% of our internal "God Mode" red team tests.
Code is up on GitHub. If you are building high-frequency trading agents or real-time voice bots, this might save you some headaches. | 2025-12-22T03:36:48 | https://github.com/rom-mvp/vigil | Fantastic-Issue1020 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1pspfft | false | null | t3_1pspfft | /r/LocalLLaMA/comments/1pspfft/i_got_tired_of_guardrails_adding_200ms_latency_to/ | false | false | default | 11 | null |
RAG Paper 25.12.18 | 5 | 1. [MEPIC: Memory Efficient Position Independent Caching for LLM Serving](http://arxiv.org/abs/2512.16822v1)
2. [Exploration of Augmentation Strategies in Multi-modal Retrieval-Augmented Generation for the Biomedical Domain: A Case Study Evaluating Question Answering in Glycobiology](http://arxiv.org/abs/2512.16802v1)
3. [From Facts to Conclusions : Integrating Deductive Reasoning in Retrieval-Augmented LLMs](http://arxiv.org/abs/2512.16795v1)
4. [DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI](http://arxiv.org/abs/2512.16676v1)
5. [Introducing ORKG ASK: an AI-driven Scholarly Literature Search and Exploration System Taking a Neuro-Symbolic Approach](http://arxiv.org/abs/2512.16425v1)
6. [Kascade: A Practical Sparse Attention Method for Long-Context LLM Inference](http://arxiv.org/abs/2512.16391v1)
7. [The Evolution of Reranking Models in Information Retrieval: From Heuristic Methods to Large Language Models](http://arxiv.org/abs/2512.16236v1)
**Collected by OpenBMB, transferred by** [**RagView.ai**](https://www.ragview.ai/) **/** [**github/RagView**](https://github.com/RagView/RagView) **.** | 2025-12-22T03:16:43 | https://www.reddit.com/r/LocalLLaMA/comments/1psp17x/rag_paper_251218/ | Cheryl_Apple | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psp17x | false | null | t3_1psp17x | /r/LocalLLaMA/comments/1psp17x/rag_paper_251218/ | false | false | self | 5 | null |
RAG Paper 25.12.15 | 1 | [removed] | 2025-12-22T03:15:15 | https://www.reddit.com/r/LocalLLaMA/comments/1psp05m/rag_paper_251215/ | Cheryl_Apple | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psp05m | false | null | t3_1psp05m | /r/LocalLLaMA/comments/1psp05m/rag_paper_251215/ | false | false | self | 1 | null |
Distributed Local LLMs | 1 | Love that we can run awesome models now but it sucks that hardware cost is getting super crazy for hobby people like us. I am working on creating a distributed network where users can join and contribute their resources so people can run awesome models like Z Image and more.
This is an MVP and planning to add more models and other cool features:
[https://picshapes.com/workers](https://picshapes.com/workers) | 2025-12-22T02:43:29 | https://www.reddit.com/r/LocalLLaMA/comments/1psocxn/distributed_local_llms/ | LiveMinute5598 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psocxn | false | null | t3_1psocxn | /r/LocalLLaMA/comments/1psocxn/distributed_local_llms/ | false | false | self | 1 | null |
MiniMax-M2 Q3_K_M on Quad V100 32gb llama.cpp testing NVlink | 22 | Almost a year ago I bought a server capable of four SXM2 GPUs. The catch was to hack the OCP power supply.
[https://www.reddit.com/r/homelab/comments/1j0769h/comment/mf8yacv/?utm\_source=share&utm\_medium=web3x&utm\_name=web3xcss&utm\_term=1&utm\_content=share\_button](https://www.reddit.com/r/homelab/comments/1j0769h/comment/mf8yacv/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button)
I actually did that properly on first attempt, but didn't torque the screws enough on the V100. It wouldn't boot. I didn't really trouble shoot further since I got busy. The project sat for a year as toyed around with Dual 5090, Quad 3090 and 12x MI50 32gb RPC. I got interested in the V100 again after seeing cheap adapters from China. Bought a boat load of 16gb adapter variants since they sold for a song and started putting together with Turbo adapters. Then with the V100 top of mind, I got four of the 32gb SXM2 and went back to the NVLink build.
tldr. Exactly as mentioned in [how do I enable NVLink / peer transfers? · ggml-org/llama.cpp · Discussion #11485 · GitHub](https://github.com/ggml-org/llama.cpp/discussions/11485), Split mode 'row' is not optimized for NVlink.
--split-mode row
About 70 tok/s pp and 20 tok/s out
slot launch_slot_: id 3 | task -1 | sampler chain: logits -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
slot launch_slot_: id 3 | task 6677 | processing task
slot update_slots: id 3 | task 6677 | new prompt, n_ctx_slot = 40192, n_keep = 0, task.n_tokens = 52
slot update_slots: id 3 | task 6677 | n_tokens = 18, memory_seq_rm [18, end)
slot update_slots: id 3 | task 6677 | prompt processing progress, n_tokens = 52, batch.n_tokens = 34, progress = 1.000000
slot update_slots: id 3 | task 6677 | prompt done, n_tokens = 52, batch.n_tokens = 34
slot print_timing: id 3 | task 6677 |
prompt eval time = 479.55 ms / 34 tokens ( 14.10 ms per token, 70.90 tokens per second)
eval time = 310990.17 ms / 6236 tokens ( 49.87 ms per token, 20.05 tokens per second)
total time = 311469.71 ms / 6270 tokens
`--split-mode layer`
Holy crap...
slot launch_slot_: id 2 | task -1 | sampler chain: logits -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
slot launch_slot_: id 2 | task 273 | processing task
slot update_slots: id 2 | task 273 | new prompt, n_ctx_slot = 40192, n_keep = 0, task.n_tokens = 52
slot update_slots: id 2 | task 273 | n_tokens = 15, memory_seq_rm [15, end)
slot update_slots: id 2 | task 273 | prompt processing progress, n_tokens = 52, batch.n_tokens = 37, progress = 1.000000
slot update_slots: id 2 | task 273 | prompt done, n_tokens = 52, batch.n_tokens = 37
slot print_timing: id 2 | task 273 |
prompt eval time = 21.97 ms / 37 tokens ( 0.59 ms per token, 1683.88 tokens per second)
eval time = 167754.38 ms / 6476 tokens ( 25.90 ms per token, 38.60 tokens per second)
total time = 167776.36 ms / 6513 tokens
Hope one day someone decides to optimize NVlink for inference. Unless you plan to train, stick with the RTX 3090 as the SXM2 systems are still highly inflated.
But consider messing with $100 V100 16gb SXM2 with a $50 adapter if you can hack cooling, or a $170 turbo adapter if you want the 5min DIY to assemble.
| 2025-12-22T02:06:37 | https://www.reddit.com/r/LocalLLaMA/comments/1psnlm0/minimaxm2_q3_k_m_on_quad_v100_32gb_llamacpp/ | MachineZer0 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psnlm0 | false | null | t3_1psnlm0 | /r/LocalLLaMA/comments/1psnlm0/minimaxm2_q3_k_m_on_quad_v100_32gb_llamacpp/ | false | false | self | 22 | {'enabled': False, 'images': [{'id': 'mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=108&crop=smart&auto=webp&s=e176d1638fb385976cf4b1d890713bee7213ed0e', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=216&crop=smart&auto=webp&s=e7b214cc2e65b9d0c5361dc29f016429518857b3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=320&crop=smart&auto=webp&s=0fef492028082b2b348bc790c99fc2ada8f1c077', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=640&crop=smart&auto=webp&s=3d3ed8bf88758207e554b56f4937db5589d0ee12', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=960&crop=smart&auto=webp&s=c22cc1aa863dfc01fa16d2dce80abd0621cfbe83', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?width=1080&crop=smart&auto=webp&s=3d93a467132acbcaad6870d5597feb0a347759c0', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/mOKSYcwU5xdC76OovqNuEuM5bp8pNHwr9WlQmBhUiJA.png?auto=webp&s=99ac03f02ca5a03817bde260fb4ce2bbc580198c', 'width': 1200}, 'variants': {}}]} |
who is Zebra on Lmarena | 0 | ive try a battle of code on Lm arena and I have got awesome result with an ai. after the vote the website called it Zebra. I have tried to ask to add an author link to the code but they both pretend to be ClaudeAi. one of them was Chat GPT and the other Zebra.
the website is broken right now...
Any idea? | 2025-12-22T01:41:56 | https://www.reddit.com/r/LocalLLaMA/comments/1psn3bm/who_is_zebra_on_lmarena/ | Ok_Contribution8157 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psn3bm | false | null | t3_1psn3bm | /r/LocalLLaMA/comments/1psn3bm/who_is_zebra_on_lmarena/ | false | false | self | 0 | null |
Revibe is a Rust-rewrite of Mistral Vibe written by Devstral 2 | 50 | Hi r/locallama,
This was my project to evaluate Devstral 2 since it's free right now. Overall, I thought it did pretty well! The CLI it made is totally usable and has a bit better performance than the original when actively agenting (not that it really matters since it'll likely be dwarfed by the model). I usually prefer tools like this to be in rust though since it's the language I work in daily.
Unfortunately, the 120b devstral is too big & slow for my hardware, but I might try to finetune the 24b. I hope Mistral and other labs will continue releasing open code models :) | 2025-12-22T00:48:25 | https://github.com/nicksenger/revibe | biet_roi | github.com | 1970-01-01T00:00:00 | 0 | {} | 1pslzv6 | false | null | t3_1pslzv6 | /r/LocalLLaMA/comments/1pslzv6/revibe_is_a_rustrewrite_of_mistral_vibe_written/ | false | false | default | 50 | {'enabled': False, 'images': [{'id': 'J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=108&crop=smart&auto=webp&s=9b2c178fdbac5b44b1e61ce7d32e6e60b58312c4', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=216&crop=smart&auto=webp&s=1cbe46ac4a1e7b20fd460f1c44e52a240bf9b063', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=320&crop=smart&auto=webp&s=10deb230021df5646879827b9c957c603c3dab45', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=640&crop=smart&auto=webp&s=c7b7531b384cf525c3c5142217cba6a53acbdba0', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=960&crop=smart&auto=webp&s=7e237a6f91cf7863931018eb795e73bdb5d9e440', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?width=1080&crop=smart&auto=webp&s=ed9772ad27836eca1a8f0d31247143947d48d4da', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/J8tu8oCcj88QnyOGI0-nxwGZljaEo8sk3VorQPTha8k.png?auto=webp&s=09ff82cb930d61a52b2450ca7a8a0c798b8a7869', 'width': 1200}, 'variants': {}}]} |
Test footage: Meta’s SAM Audio guitar isolation | 4 | Posting test footage from ongoing work building an API around Meta’s SAM Audio.
This is a comparison using the same source clip processed with the Small, Base, and Large models for guitar isolation. The Large model is predictably the cleanest, with less bleed and better transient detail, while Small struggles. Base sits somewhere in between.
Inference was run on an NVIDIA H200.
**- Source:** [https://drive.google.com/file/d/1mLwq1CJ\_6NAwaChOX85By-Ud\_ULdF7Or/view?usp=sharing](https://drive.google.com/file/d/1mLwq1CJ_6NAwaChOX85By-Ud_ULdF7Or/view?usp=sharing)
**- SAM Large:** [https://drive.google.com/file/d/187GGmgjiyZ5G2seClPb-iyt2LM0cTdow/view?usp=sharing](https://drive.google.com/file/d/187GGmgjiyZ5G2seClPb-iyt2LM0cTdow/view?usp=sharing)
**- SAM Base:** [https://drive.google.com/file/d/1RaY49Gf561DYGgmOZ6ULCJHfVzBw2DMw/view?usp=sharing](https://drive.google.com/file/d/1RaY49Gf561DYGgmOZ6ULCJHfVzBw2DMw/view?usp=sharing)
**- SAM Small:** [https://drive.google.com/file/d/1DSiyTuZe0Qf3U9TLCPu08EgxhgrcwXa9/view?usp=sharing](https://drive.google.com/file/d/1DSiyTuZe0Qf3U9TLCPu08EgxhgrcwXa9/view?usp=sharing)
| 2025-12-22T00:42:29 | https://drive.google.com/file/d/1mLwq1CJ_6NAwaChOX85By-Ud_ULdF7Or/view?usp=sharing | pzzle-nj | drive.google.com | 1970-01-01T00:00:00 | 0 | {} | 1pslvje | false | null | t3_1pslvje | /r/LocalLLaMA/comments/1pslvje/test_footage_metas_sam_audio_guitar_isolation/ | false | false | default | 4 | null |
My experience quiet cooling 2 external/open-air Instinct MI50 cards. | 13 | Just FYI for anyone wanting to quietly cool their MI50 cards.
**Background**
With the stock MI50 cover/radiators, I would expect you will get best results with a blower-type fan. Since my cards are external, I have plenty of room, so wanted to go with 120mm blowers. On Ebay I could only find 80 mm blowers with shrouds, but wanted to go bigger for quieter cooling. Apparently there's not a big market for blowers designed to be quiet, really only found 1: the [AC Infinity MULTIFAN S2](https://www.amazon.com/dp/B012CL2V3I). I also ordered a Wathal fan that was much louder, but much more powerful, but unnecessary.
The AC Infinity fan is powered by USB, so I have it plugged into the USB outlet on my server (A Minisforum MS-A2). This is kinda nice since it turns the fans on and off with the computer, but what I may do is see if I can kill power to the USB ports, monitor the cards temps, and only power the fans when needed (there are commands that are supposed to be able to do this, but haven't tried on my hardware, yet).
**Results**
Using AC Infinity MULTIFAN S2 on _lowest setting_, maxing it out with llama-bench sustained load with 8K prompt through 100 repititions, maxes out and stays at **70-75 C**. The rated max for MI50 is 94 C but want to keep 10-15 lower than max under load, which this manages no problem. On highest fan setting, keeps it about 60 C and is still pretty quiet. Lowest fan setting drops it back down pretty quick to 30 C, takes a long time to get it up to 75 C going from idle to maxed out.
Here is the exact command I ran (I ran it twice to get 100 (killed the first run when it started TG testing:
./llama-bench -m ~/.cache/llama.cpp/unsloth_Qwen3-Next-80B-A3B-Instruct-GGUF_Qwen3-Next-80B-A3B-Instruct-UD-Q4_K_XL.gguf -sm layer -fa 1 --cache-type-k q8_0 --cache-type-v q8_0 --progress -p 8192 -n 128 -r 100
I've done a ton of testing on what models can run at speeds I'm comfortable with, and this pretty closely mimics what I'm planning to run with llama-server indefinitely, although it will be mostly idle and will not run sustained inference for anywhere near this duration.
It took 13 minutes (prompt run 55) to reach 75 C. It gets up to 55 C after a minute or 2 and then creeps up slower and slower. The absolute highest temp I saw (using "sudo rocm-smi --alldevices --showtempgraph") was 76 C; it mostly bounced around 72 - 74 C.
**Caveats**
Probably the biggest thing to consider is that the model is running split between 2 cards. A model running on a single card may keep that single card more sustained at maximum load.
Um... I guess that's the only caveat I can think of right now.
**Power**
Additional FYI - I'm running both cards off a single external PSU with splitter cables, connected to a watt-meter, most power draw I'm seeing is 250W. I didn't set any power limiting. So this also supports the caveat that something
Idle power draw for both cards was consistently 38 W.
**Attaching The Fans**
I just used blue painter's tape. | 2025-12-22T00:09:13 | https://www.reddit.com/r/LocalLLaMA/comments/1psl6nd/my_experience_quiet_cooling_2_externalopenair/ | moderately-extremist | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psl6nd | false | null | t3_1psl6nd | /r/LocalLLaMA/comments/1psl6nd/my_experience_quiet_cooling_2_externalopenair/ | false | false | self | 13 | {'enabled': False, 'images': [{'id': 'wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=108&crop=smart&auto=webp&s=c7ef9713fb4fbf51d0d7da30fb558f95324a395b', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=216&crop=smart&auto=webp&s=70f4ef0366eafa569960666b4537977954dc4da4', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=320&crop=smart&auto=webp&s=e88e6f574ea2b6abf3644be5140a1ed8ad6d613c', 'width': 320}, {'height': 335, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=640&crop=smart&auto=webp&s=290ace7209dd3df0a237ec970a6a8b1662d523e1', 'width': 640}, {'height': 502, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=960&crop=smart&auto=webp&s=421952297faebb04d1038184216c053ab1f0bb56', 'width': 960}, {'height': 565, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?width=1080&crop=smart&auto=webp&s=2e3704dd3e397c6dbebe004c6cce33e8cd82d316', 'width': 1080}], 'source': {'height': 1000, 'url': 'https://external-preview.redd.it/wyRlnnC4nIHWRfWMUIBnHvHMsP98N9mROJtKXbnwKWI.png?auto=webp&s=8cdb17f0919f23f3fc3c0bd9dac21cd40118adda', 'width': 1910}, 'variants': {}}]} |
Ollama benchmarks on Ryzen 5600 + RTX 3050 (8GB) | 0 | Ahoy! I’ve seen a lot of posts lately benchmarking models on massive setups (dual RTX 3090s, H100s, etc.), so I wanted to see how things look on low-to-mid consumer hardware instead.
Honestly, I’m pretty impressed by how well even larger models run on a consumer machine. The system is heavily overclocked, but that shouldn’t make a huge difference overall.
All tests were run with a 4096 context length due to RAM and VRAM limits. The evaluations are handled automatically via a Python script, and the results are exported to a CSV file, then formated by me. (Why does making a table on reddit is such a pain in the \*ss)
**Specs**
* CPU: Ryzen 5 5600 @ 4.6 GHz (OC)
* RAM: DDR4 3400 MHz CL14 (OC)
* GPU: RTX 3050 8GB (+300 core / +2000 mem)
|Model|Eval rate (tok/s)|Prompt eval|Prompt rate |Eval duration|Load|
|:-|:-|:-|:-|:-|:-|
|mistral-small3.2:24b-instruct-q4\_K\_M|4.53|5.75s|90.06|4.63s|8.00s|
|gemma3n:e4b-it-q4\_K\_M|28.25|69ms|289.73|1.56s|8.11s|
|gemma3:27b-it-qat|3.67|1.49s|24.77|12.25s|29.04s|
|gemma3:12b-it-qat|13.17|123ms|300.93|2.96s|20.21s|
|ministral-3:8b-instruct-q4\_K\_M|32.56|412ms|1369.88|1.01s|5.38s|
|ministral-3:14b-instruct-q4\_K\_M|12.64|1.22s|462.54|1.90s|6.74s|
|phi4:latest|11.37|513ms|40.94|2.29s|16.29s|
|GLM-4.6V-Flash GGUF IQ4\_NL|**59.89**|48ms|354.31|3.31s|22.77s|
|qwen3:4b-instruct-q4\_K\_M|**74.79**|29ms|649.86|455ms|**3.30s**|
|devstral-small-2:24b-q4\_K\_M|4.96|3.39s|166.78|4.83s|9.69s|
|nemotron-3-nano:30b-q4\_K\_M|15.30|669ms|40.35|4.25s|20.54s|
|qwen3-vl:30b-q4\_K\_M|26.56|332ms|57.29|1.47s|13.01s|
|qwen3-coder:30b-q4\_K\_M|27.29|292ms|65.09|1.06s|11.08s|
|gpt-oss:20b|17.11|825ms|94.56|7.54s|9.93s|
|GLM-4.6V GGUF IQ1\_M|3.45|**29.3s**|0.58|**2m30s**|240ms|
| 2025-12-21T23:55:20 | https://www.reddit.com/r/LocalLLaMA/comments/1pskvnj/ollama_benchmarks_on_ryzen_5600_rtx_3050_8gb/ | cride20 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pskvnj | false | null | t3_1pskvnj | /r/LocalLLaMA/comments/1pskvnj/ollama_benchmarks_on_ryzen_5600_rtx_3050_8gb/ | false | false | self | 0 | null |
Release Ada v3.0.0 - The Pair Programmer · luna-system/ada | 1 | [removed] | 2025-12-21T23:35:16 | https://github.com/luna-system/ada/releases/tag/v3.0.0 | dual-moon | github.com | 1970-01-01T00:00:00 | 0 | {} | 1pskgcs | false | null | t3_1pskgcs | /r/LocalLLaMA/comments/1pskgcs/release_ada_v300_the_pair_programmer_lunasystemada/ | false | false | default | 1 | null |
Ultima 2 Challenge: COMPLETED. ✅ You asked for a tile-based RPG engine with state management. The Agent delivered. | 0 | **Under the hood (as seen in the video):**
* **State Machine:** Fully implemented. Seamless switching between `OVERWORLD` and `TOWN` states based on tile triggers.
* **Persistence:** The agent handles coordinate resets when entering/exiting zones.
* **Tile Engine:** Dynamic rendering of 4 different terrain types + walls.
* **Logic:** Turn-based movement, collision detection (water/walls), and NPC interaction logic.
**Verdict:** This required maintaining context across multiple class structures and game loops. A massive win for local 30B models. | 2025-12-21T23:10:31 | https://v.redd.it/farm5azu2n8g1 | Alone-Competition863 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psjwo8 | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/farm5azu2n8g1/DASHPlaylist.mpd?a=1768950646%2CYTQzYzRmZWFhMTliNDJhMjAwNTYwYzQ4M2I3NTNlOGQ2MDI4OTNiYTZiMzJiMWY5N2NjM2NlOGVjNGFjYWMxMA%3D%3D&v=1&f=sd', 'duration': 89, 'fallback_url': 'https://v.redd.it/farm5azu2n8g1/CMAF_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/farm5azu2n8g1/HLSPlaylist.m3u8?a=1768950646%2CMjdkYTAzNjc3MTA0MWEyYjIwMjEzYjMzNWI5ZjdjODBjMTMyMmU3Mjk5Y2E4NjE0ODA4OTM1MThkYzhhYjc0NA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/farm5azu2n8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1psjwo8 | /r/LocalLLaMA/comments/1psjwo8/ultima_2_challenge_completed_you_asked_for_a/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=108&crop=smart&format=pjpg&auto=webp&s=00272b6a1dc9628a236c02445e66ff783c65670d', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=216&crop=smart&format=pjpg&auto=webp&s=3f1f70d1fbfaeaf698bf5bcebe027279f4a4a66f', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=320&crop=smart&format=pjpg&auto=webp&s=a15b6903daee82bef007670e009dcf9f31277ec8', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=640&crop=smart&format=pjpg&auto=webp&s=2b075ea23262981c4d53b821e40b4a481e3fa3c2', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=960&crop=smart&format=pjpg&auto=webp&s=c2bd2b52d40ff1d3ca7b35f2aa21db92b5904b62', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?width=1080&crop=smart&format=pjpg&auto=webp&s=7299587bb1f217880d0ef69b7dab90fd59f6300b', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/amh2YjFwM3YybjhnMf3CgSBkjdJpI-IDVjUSNCAXZBQqMepkGuNeLK6QbT2A.png?format=pjpg&auto=webp&s=2eba44909996cb96b8b69f79605cf24d84eb8941', 'width': 1920}, 'variants': {}}]} | |
Any regrets A6000 Pro owners? | 26 | Seems like quite a future proof option at the moment for local ai needs. For those who bought - how do you guys feel about your decision. What's been working for you best and where things fall short of your expectations.
I use to have threadripper rig with 6 RTX 3090, it was messy and very power hungry, then started feeling dated. Also Claude was so much better than local models, it was hard to keep as much workloads local as I wish. Now the models yet again improved and some are more decent for agentic use than before, thinking of a more clean and modern setup. | 2025-12-21T22:32:15 | https://www.reddit.com/r/LocalLLaMA/comments/1psj1pr/any_regrets_a6000_pro_owners/ | val_in_tech | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psj1pr | false | null | t3_1psj1pr | /r/LocalLLaMA/comments/1psj1pr/any_regrets_a6000_pro_owners/ | false | false | self | 26 | null |
Anyone else constantly experimenting with AI tools? | 1 | I’ve been spending way too much time testing different AI tools and models and realized it’d be way better to do it with other people.
A few of us started a small Discord to share tools, prompts, experiments, and what’s actually worth using.
If that sounds like your thing, feel free to join:
[https://discord.gg/SdX5STB6HE](https://discord.gg/SdX5STB6HE) | 2025-12-21T22:28:55 | https://www.reddit.com/r/LocalLLaMA/comments/1psiz3l/anyone_else_constantly_experimenting_with_ai_tools/ | GPTCodeShopper | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psiz3l | false | null | t3_1psiz3l | /r/LocalLLaMA/comments/1psiz3l/anyone_else_constantly_experimenting_with_ai_tools/ | false | false | self | 1 | null |
Any of these models that I can wipe? | 0 | I use a CPU with 8-ish cores and 16 GB ram, no GPU cuz I'm kinda broke. I'm not sure if this a good place to post, but are there any good models I should add to my collection blow. Anything to remove?
NAME ID SIZE MODIFIED
bge-m3:latest 790764642607 1.2 GB 9 days ago
granite-embedding:278m 1a37926bf842 562 MB 9 days ago
hf.co/bartowski/allenai_Olmo-3-7B-Think-GGUF:Q4_K_M af357bd95bbc 4.5 GB 2 weeks ago
hf.co/gabriellarson/Moonlight-16B-A3B-Instruct-GGUF:Q4_K_M b63f8b621e59 10 GB 3 weeks ago
hf.co/LiquidAI/LFM2-8B-A1B-GGUF:Q4_K_M 5b905e4f6020 5.0 GB 3 weeks ago
hf.co/allenai/OLMoE-1B-7B-0125-GGUF:Q4_K_M e6410246d2a5 4.2 GB 3 weeks ago
ibm/granite3.2-guardian:3b 4f7cdff62b7e 2.7 GB 4 weeks ago
ibm/granite3.3-guardian:8b 90a8aabc98eb 6.7 GB 4 weeks ago
smollm2:1.7b cef4a1e09247 1.8 GB 4 weeks ago
qwen2.5-coder:latest dae161e27b0e 4.7 GB 4 weeks ago
deepseek-r1:7b 755ced02ce7b 4.7 GB 4 weeks ago
deepseek-r1:1.5b e0979632db5a 1.1 GB 4 weeks ago
granite3.1-moe:3b b43d80d7fca7 2.0 GB 4 weeks ago
NeuroEquality/safe4all-guardian:latest 0a467f0d2db9 1.8 GB 4 weeks ago
ibm/granite3.3-vision:2b 5cdfea23a292 3.6 GB 4 weeks ago
dolphin-phi:2.7b c5761fc77240 1.6 GB 4 weeks ago
qwen2.5-coder:7b dae161e27b0e 4.7 GB 4 weeks ago
command-r7b:latest ff4e9696ef9f 5.1 GB 4 weeks ago
hf.co/adamo1139/Danube3-4b-4chan-HESOYAM-2510-GGUF:Q8_0 d4946d50a87c 4.2 GB 5 weeks ago
dolphin-mistral:7b 5dc8c5a2be65 4.1 GB 5 weeks ago
dolphin-mistral:latest 5dc8c5a2be65 4.1 GB 5 weeks ago
dolphin-phi:latest c5761fc77240 1.6 GB 5 weeks ago
granite4:tiny-h c4d3ac2a16a3 4.2 GB 7 weeks ago
qwen2.5-coder:1.5b d7372fd82851 986 MB 7 weeks ago
qwen2.5-coder:3b f72c60cabf62 1.9 GB 7 weeks ago
artifish/llama3.2-uncensored:latest c73bea26e004 2.2 GB 7 weeks ago
qwen3:4b 359d7dd4bcda 2.5 GB 7 weeks ago
I know this may come off as low effort or shit-posty but I really am not trying to do that. I have too many models and need to purge some while adding some newbies. | 2025-12-21T22:23:11 | https://www.reddit.com/r/LocalLLaMA/comments/1psiui4/any_of_these_models_that_i_can_wipe/ | Brospeh-Stalin | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psiui4 | false | null | t3_1psiui4 | /r/LocalLLaMA/comments/1psiui4/any_of_these_models_that_i_can_wipe/ | false | false | self | 0 | null |
Using local VLMs and SAM 3 to Agentically Segment Characters | 18 | It's been my goal for a while to come up with a reliable way to segment characters in an automated way, (hence why I built my [Sa2VA node](https://github.com/adambarbato/ComfyUI-Sa2VA)), so I was excited when SAM 3 released last month. Just like its predecessor, SAM 3 is great at segmenting the general concepts it knows and is even better than SAM 2 and can do simple noun phrases like "blonde woman". However, that's not good enough for character-specific segmentation descriptions like "the fourth woman from the left holding a suitcase".
But at the same time that SAM 3 released, I started hearing people talk about the [SAM 3 Agent](https://github.com/facebookresearch/sam3/blob/main/examples/sam3_agent.ipynb) example notebook that the authors released showing how SAM 3 could be used in an agentic workflow with a VLM. I wanted to put that to the test, so [I adapted their workbook into a ComfyUI node](https://github.com/adambarbato/ComfyUI-Segmentation-Agent) that works with both local GGUF VLMs (via llama-cpp-python) and through OpenRouter.
## How It Works
1. The agent analyzes the base image and character description prompt
2. It chooses one or more appropriate simple noun phrases for segmentation (e.g., "woman", "brown hair", "red dress") that will likely be known by the SAM 3 model
3. SAM 3 generates masks for those phrases
4. The masks are numbered and visualized on the original image and shown to the agent
5. The agent evaluates if the masks correctly segment the character
6. If correct, it accepts all or a subset of the masks that best cover the intended character; if not, it tries additional phrases
7. This iterates until satisfactory masks are found or max_iterations is reached and the agent fails
### Limitations
This agentic process works, but the results are often worse (and much slower) than purpose-trained solutions like [Grounded SAM](https://github.com/IDEA-Research/Grounded-SAM-2) and [Sa2VA](https://github.com/bytedance/Sa2VA). The agentic method CAN get even more correct results than those solutions if used with frontier vision models (mostly the Gemini series from Google) but I've found that the rate of hallucinations from the VLM often cancels out the benefits of checking the segmentation results rather than going with the 1-shot approach of Grounded SAM/Sa2VA.
This may still be the best approach if your use case needs to be 100% agentic and can tolerate long latencies and needs the absolute highest accuracy. I suspect using frontier VLMs paired with many more iterations and a more aggressive system prompt may increase accuracy at the cost of price and speed.
Personally though, I think I'm sticking to Sa2VA for now for its good-enough segmentation and fast speed.
### Future Improvements
1. Refine the system prompt to include known-good SAM 3 prompts
- A lot of the system's current slowness involves the first few steps where the agent may try phrases that are too complicated for SAM and result in 0 masks being generated (often this is just a rephrasing of the user's initial prompt). Including a larger list of known-useful SAM 3 prompts may help speed up the agentic loop at the cost of more system prompt tokens.
2. Use the same agentic loop but with Grounded SAM or Sa2VA
- What may produce the best results is to pair this agentic loop with one of the segmentation solutions that has a more open vocabulary. Although not as powerful as the new SAM 3, Grounded SAM or Sa2VA may play better with the verbose tendencies of most VLMs and their smaller number of masks produced per prompt may help cut down on hallucinations.
3. Try with bounding box/pointing VLMs like Moondream
- The original SAM 3 Agent (which is reproduced here) uses text prompts from the VLM to SAM to indicate what should be segmented, but, as mentioned, SAM's native language is not text, it's visuals. Some VLMs (like the Moondream series) are trained to produce bounding boxes/points. Putting one of those into a similar agentic loop may reduce the issues described above, but may introduce its own issue in deciding what each system considers segmentable within a bounding box.
### Quick Links
- GitHub Repo: https://github.com/adambarbato/ComfyUI-Segmentation-Agent
- Example ComfyUI workflow: https://github.com/adambarbato/ComfyUI-Segmentation-Agent/blob/main/workflow/comfyui-segment-agent.json | 2025-12-21T22:17:57 | Complete-Lawfulness | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psiq8a | false | null | t3_1psiq8a | /r/LocalLLaMA/comments/1psiq8a/using_local_vlms_and_sam_3_to_agentically_segment/ | false | false | default | 18 | {'enabled': True, 'images': [{'id': 'z1a1qt5otm8g1', 'resolutions': [{'height': 51, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=108&crop=smart&auto=webp&s=d4d49efd558b6b000742a8e75414cf2585680351', 'width': 108}, {'height': 103, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=216&crop=smart&auto=webp&s=b980155ee1dd2cc58cb7d5675de1407b873b1ead', 'width': 216}, {'height': 153, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=320&crop=smart&auto=webp&s=15afab28ae68ae593f13a71db3a41f7aaddf2970', 'width': 320}, {'height': 307, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=640&crop=smart&auto=webp&s=f37a6822173e8ed9be99727ccdf9a798e0ffc9a4', 'width': 640}, {'height': 460, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=960&crop=smart&auto=webp&s=a4d3a1d1e4aadc78cadcc5e049ab83f97fe98552', 'width': 960}, {'height': 518, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?width=1080&crop=smart&auto=webp&s=8cddd51f5dcb597755de3e66f89f8fce9a8c24ec', 'width': 1080}], 'source': {'height': 1078, 'url': 'https://preview.redd.it/z1a1qt5otm8g1.png?auto=webp&s=0d53242085129000b6ff26258df85ccf8e2af979', 'width': 2246}, 'variants': {}}]} | |
What is the best model for coding in local 8-14b parameters | 5 | I saw nvidia nemotron cascade etc. Results but im not sure they are working really well. Im not looking for sonnet grade model. I plan to fine-tune it for cyber security tasks. And also i have 300 dollar google cloud credit but i cant use it in gpu. Can i finetune with gpu? | 2025-12-21T21:59:46 | https://www.reddit.com/r/LocalLLaMA/comments/1psib5j/what_is_the_best_model_for_coding_in_local_814b/ | nicklazimbana | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psib5j | false | null | t3_1psib5j | /r/LocalLLaMA/comments/1psib5j/what_is_the_best_model_for_coding_in_local_814b/ | false | false | self | 5 | null |
I built an MCP server for stock analysis (79% val. accuracy) – Ensemble LSTM/RL model accessible via natural language | 0 | ```markdown
I've been working on a project to bridge quantitative finance models with LLMs using the Model Context Protocol (MCP).
I just released **InvestBuddy**, an MCP server that connects LLMs (currently optimized for Claude Desktop, but technically compatible with any MCP client) to a custom ensemble model I built.
## The Architecture
**Ensemble ML:** Combines LSTM (for sequence prediction) + Reinforcement Learning (for portfolio optimization) + Transformers (for sentiment).
**Model Tag:** `v20251130_correlation`
**Validation:** Backtested on 12,901 predictions (S&P 100) with a 2-year walk-forward window (2023-2025).
**Stats:**
- Sharpe Ratio: 2.34
- Directional accuracy: ~79% on validation set
- Statistical significance: p < 0.000001 (t-stat: 28.45)
- Full methodology: [investbuddy.ai/transparency](https://www.investbuddy.ai/transparency)
## What it exposes to the LLM
The MCP server provides 5 tools:
1. **`get_stock_prediction`** – 10-day price forecasts with confidence intervals
2. **`get_market_regime`** – Detects Bull/Bear/Sideways trends using HMM
3. **`analyze_portfolio`** – Returns optimal weights based on risk tolerance (RL-based)
4. **`discover_stocks`** – AI screening for undervalued/momentum opportunities
5. **`batch_predict`** – Parallel predictions for multiple tickers
## Why I'm sharing here
I know this sub focuses on local models, but I think **MCP is a crucial layer** for making agents (local or hosted) actually useful. This server allows an LLM to "outsource" the heavy math to a specialized ML model rather than hallucinating numbers.
The LLM handles natural language parsing, the finance model handles quantitative prediction. Clean separation of concerns.
## Try it out
- **Smithery listing:** [smithery.ai/server/investbuddy-mcp/investbuddy-mcp-server](https://smithery.ai/server/investbuddy-mcp/investbuddy-mcp-server)
- **Direct install:** `npx -y investbuddy-mcp-server`
- **GitHub:** [github.com/danielrz/investbuddy-mcp-server](https://github.com/danielrz/investbuddy-mcp-server)
**Access:** There is a free tier (5 calls/day) so you can test the accuracy without paying. Documentation is at [investbuddy.ai/mcp](https://investbuddy.ai/mcp). | 2025-12-21T21:54:10 | https://www.reddit.com/gallery/1psi6iu | InvestBuddy_AI | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1psi6iu | false | null | t3_1psi6iu | /r/LocalLLaMA/comments/1psi6iu/i_built_an_mcp_server_for_stock_analysis_79_val/ | false | false | 0 | null | |
When is Anthropic going to release a 120b for the community? Are they scared they can't beat OpenAI? | 0 | Where is it? :) | 2025-12-21T21:52:48 | https://www.reddit.com/r/LocalLLaMA/comments/1psi5hq/when_is_anthropic_going_to_release_a_120b_for_the/ | Aggressive-Bother470 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psi5hq | false | null | t3_1psi5hq | /r/LocalLLaMA/comments/1psi5hq/when_is_anthropic_going_to_release_a_120b_for_the/ | false | false | self | 0 | null |
Building an Autonomous "AI Auditor" for ISO Compliance: How would you architect this for production? | 0 | I am building an agentic workflow to automate the documentation review process for third-party certification bodies. I have already built a functional prototype using Google Anti-gravity based on a DOE (Directive-Orchestration-Execution) framework, but now I need to determine the absolute best stack to rebuild this for a robust, enterprise-grade production environment.
The Business Process (The "What"):
Ingestion: The system receives a ZIP file containing complex unstructured audit evidence (PDFs, images, technical drawings, scanned hand-written notes).
Context Recognition: It identifies the applicable ISO standard (e.g., 9001, 27001) and any integrated schemes.
Dynamic Retrieval: It retrieves the specific Audit Protocols and SOPs for that exact standard from a knowledge base.
Multimodal Analysis (The Core): Instead of using brittle OCR/Python text extraction scripts, I am leveraging Gemini 1.5/3 Pro’s multimodal capabilities to visually analyze the evidence, "see" the context, and cross-reference it against the ISO clauses.
Output Generation: The agent must perfectly fill out a rigid, complex compliance checklist (Excel/JSON) and flag specific non-conformities for the human auditor to review.
The Challenge:
The prototype proves the logic works, but moving from a notebook environment to a production system that processes massive files without crashing is a different beast.
My Questions for the Community:
Orchestration & State: For a workflow this heavy (long-running processes, handling large ZIPs, multiple reasoning steps per document), what architecture do you swear by to manage state and handle retries? I need something that won't fail if an API hangs for 30 seconds.
Structured Integrity: The output checklists must be 100% syntactically correct to map into legacy Excel files. What is the current "gold standard" approach for forcing strictly formatted schemas from multimodal LLM inputs without degrading the reasoning quality?
RAG Strategy for Compliance: ISO standards are hierarchical and cross-referenced. How would you structure the retrieval system (DB type, indexing strategy) to ensure the agent pulls the exact clause it needs, rather than just generic semantic matches?
Goal: I want a system that is anti-fragile, deterministic, and scalable. How would you build this today? | 2025-12-21T21:51:46 | https://www.reddit.com/r/LocalLLaMA/comments/1psi4ot/building_an_autonomous_ai_auditor_for_iso/ | doctorallfix | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psi4ot | false | null | t3_1psi4ot | /r/LocalLLaMA/comments/1psi4ot/building_an_autonomous_ai_auditor_for_iso/ | false | false | self | 0 | null |
Trying to understand benchmarks | 0 | I’m new to this but from some posts and benchmarks it seems that people are saying that gpt-oss-20B (high) is smarter that 4o.
Does this mean that the model I run locally is better than the model I used to pay for monthly?
What am I misunderstanding? | 2025-12-21T21:50:03 | https://www.reddit.com/r/LocalLLaMA/comments/1psi3ap/trying_to_understand_benchmarks/ | butt_badg3r | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psi3ap | false | null | t3_1psi3ap | /r/LocalLLaMA/comments/1psi3ap/trying_to_understand_benchmarks/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=108&crop=smart&auto=webp&s=700f91dbca11e5a7030b915550ae877ef725a0d4', 'width': 108}, {'height': 120, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=216&crop=smart&auto=webp&s=b97954336b79c1390848d0e44fa056a85de68672', 'width': 216}, {'height': 177, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=320&crop=smart&auto=webp&s=65f53b80ab9674ee645013e3e8eeac4f953d657e', 'width': 320}, {'height': 355, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=640&crop=smart&auto=webp&s=47f397e4a22ed5ec7e82aad070eb446319603abc', 'width': 640}, {'height': 533, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=960&crop=smart&auto=webp&s=0f4359d47b78f5c1aa35de8804dbe36a749fc11a', 'width': 960}, {'height': 600, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?width=1080&crop=smart&auto=webp&s=62eb4b7216f41af6600fc4df79cfa67425c19442', 'width': 1080}], 'source': {'height': 700, 'url': 'https://external-preview.redd.it/RVufpUx3tddh_BCVq7ZzBCU7nLRDZ_d0EprIuvN6J-E.png?auto=webp&s=efc17c9f241b4403d22cbacfe5d71900ee1cf85a', 'width': 1260}, 'variants': {}}]} |
Looking for a local LLM that can help me understand a very large, complex proprietary codebase | 2 | Hey everyone,
I’ve recently started experimenting with **local LLMs** (Ollama ecosystem), so please excuse any beginner mistakes.
I’ve already Googled around and tried the commonly suggested setups, but I’m hitting real limitations and would appreciate guidance from people who’ve done this successfully.
Situation is that I recently started a new job and inherited a **very large proprietary system.** This system consist of:
* \~130 projects in a single solution
* A few UI projects (Angular being main one, but there are others)
* Only 2 out of \~30 developers truly understand how the system works
* The system is **not documented at all**
* I cannot upload code to cloud LLMs for IP reasons
Because of this, I’m trying to use **local LLMs** to:
* Ask architectural questions
* Trace data flow
* Understand how UI is dynamically rendered
* Get explanations *based strictly on the existing code*
**My Hardware is below:**
* RTX 4070 SUPER
* 32 GB DDR5 6000 MHz
* Ryzen 7600X
**Models (via Ollama)**
* `qwen3-coder:30b`
* `qwen3-coder-30b-q5`
* `qwen3:30b`
**Tooling**
* VS Code + Continue extension
* I tried using "continue" VS code extension, but it lacks context (or adding context is freaking hard) so I abandoned it.
* VS Code + GitHub Copilot (local models)
* I found I can use the GitHub copilot in VS Code with local models so I started using it, mainly due to the `@workspace` tag. However, this is not yielding any results. Model is literally making stuff up even though it takes over 70 references.
* Literally says it found something which is not there in project at all.
#
My main issue is that even when the model claims to reference **dozens of files**, it hallucinates components that **do not exist.** Also, it claims functionality that is **nowhere in the codebase.**
Best results I got is when it starts explanations correctly, then derails halfway
This happens even for **very concrete questions** like:
>
**To give some more context how I use it:**
As stated above, one project is written in Angular - with whom I never worked with.
This Angular app pulls **HTML input definitions + CSS** from the database and renders them dynamically. (I mean like literal HTML input elements with css alongside them).
I open this folder where Angular project is VS code and basically ask "You are senior Angular dev bla bla bla ... Find me example and explain to me how does this dynamic rendering of UI elements work.
My question is:
*Is this fundamentally a* ***model limitation****, or am I using the wrong approach/tools?*
Specifically:
* Is there a **local model** that is better at *grounded code understanding* for very large codebases?
* Is there a better **workflow** than VS Code + Continue / Copilot for this use case?
* Should I be chunking/indexing the project differently (RAG, embeddings, etc.)?
* Or is expecting accurate reasoning over a 130-project solution unrealistic with today’s local models?
I’m not looking for autocomplete — I need **accurate, non-hallucinated explanations tied to real code**.
Any advice from people doing serious local LLM + large codebase analysis would be hugely appreciated.
Thanks! | 2025-12-21T21:46:47 | https://www.reddit.com/r/LocalLLaMA/comments/1psi0l9/looking_for_a_local_llm_that_can_help_me/ | only_4kids | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psi0l9 | false | null | t3_1psi0l9 | /r/LocalLLaMA/comments/1psi0l9/looking_for_a_local_llm_that_can_help_me/ | false | false | self | 2 | null |
I built a website that aggregates latest challenges in research papers as ideas for people who are looking for something challenging to work on … I hope something like this can be useful for anyone. | 11 | This is the website link:
[https://research-ideas.org/](https://research-ideas.org/)
The ideas in it gets updated daily, you can search on a specific topic you want to get some ideas in or click on random to leave the website choose an idea for you. | 2025-12-21T21:20:37 | Severe-Awareness829 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pshesg | false | null | t3_1pshesg | /r/LocalLLaMA/comments/1pshesg/i_built_a_website_that_aggregates_latest/ | false | false | default | 11 | {'enabled': True, 'images': [{'id': 'ixsxb94xim8g1', 'resolutions': [{'height': 44, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=108&crop=smart&auto=webp&s=fb49a5495f08cf54e681bafedba00ac1adf49cf8', 'width': 108}, {'height': 88, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=216&crop=smart&auto=webp&s=08a1219145aa02c3f0135dc8e4945d8f7a7dba41', 'width': 216}, {'height': 131, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=320&crop=smart&auto=webp&s=f598bbaae1d0023adb52762fc39123eabcfb378c', 'width': 320}, {'height': 263, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=640&crop=smart&auto=webp&s=8a2d67e0e2948c01c76856f214092311084a9b21', 'width': 640}, {'height': 394, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=960&crop=smart&auto=webp&s=d597c0123619e9fb3cda2426ae21ee32e087033a', 'width': 960}, {'height': 443, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?width=1080&crop=smart&auto=webp&s=cfc46fc850729f0ba7d9fd5aeee974f951ce2a83', 'width': 1080}], 'source': {'height': 734, 'url': 'https://preview.redd.it/ixsxb94xim8g1.png?auto=webp&s=40f2a0e39692307c9ba7e0677a502daada2ec64c', 'width': 1786}, 'variants': {}}]} | |
1 year later and people are still speedrunning NanoGPT. Last time this was posted the WR was 8.2 min. Its now 127.7 sec. | 209 | [Previous post](https://www.reddit.com/r/LocalLLaMA/comments/1gmd1a8/are_people_speedrunning_training_gpts_now/) for context. Also note original NanoGPT run from Andrej Karpathy was 45 min. I think this is a great way to understand progress in overall algorithmic speed improvements as I'm sure the big labs are using similar speedup tricks. | 2025-12-21T21:04:59 | jd_3d | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psh1w2 | false | null | t3_1psh1w2 | /r/LocalLLaMA/comments/1psh1w2/1_year_later_and_people_are_still_speedrunning/ | false | false | default | 209 | {'enabled': True, 'images': [{'id': 'fkz64bswfm8g1', 'resolutions': [{'height': 86, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=108&crop=smart&auto=webp&s=fc00a400d30dda2fc82c8a6f9755b8f7c5e26efa', 'width': 108}, {'height': 172, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=216&crop=smart&auto=webp&s=f32bc61fddc8070187c00a9a5b3304d81badef7c', 'width': 216}, {'height': 256, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=320&crop=smart&auto=webp&s=4ed333a39386814c0afbe2956aeb066ec7b231e3', 'width': 320}, {'height': 512, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=640&crop=smart&auto=webp&s=173d94496a9f94631434a2e6b566db19eb11ba40', 'width': 640}, {'height': 768, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=960&crop=smart&auto=webp&s=bf84f03bd75b2b09d28e47e60ae1dad5b5e67ea6', 'width': 960}, {'height': 864, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?width=1080&crop=smart&auto=webp&s=040d10554b7b50884802ecbf098c71507cfa4f59', 'width': 1080}], 'source': {'height': 1053, 'url': 'https://preview.redd.it/fkz64bswfm8g1.png?auto=webp&s=d0e1af597ce01ff9f09798222d597d3423e6a341', 'width': 1316}, 'variants': {}}]} | |
My Local Agent built this Stealth Game in one go. I’m tired of choosing projects. YOU tell me what to build next. | 0 | Running Qwen3-30B locally on RTX 4070. People think these videos are cherry-picked. Fine.
1. Watch the video (It handled raycasting, AI patrol paths, and collision logic autonomously).
2. **Comment a game idea/mechanic below.**
3. I will feed the **top upvoted comment** directly into the agent as a prompt – UNEDITED.
4. I will post the result tomorrow. | 2025-12-21T20:55:36 | https://v.redd.it/99r6pskqem8g1 | Alone-Competition863 | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psgu8b | false | {'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/99r6pskqem8g1/DASHPlaylist.mpd?a=1768942553%2CNzQzYzk0ODBlNGQzOTFlNGY1NGY1ODU4Njg1NTE3MDYyODZjYjQyNjUzZWRmOWYxMWQ3MmFkMDcwNmJjZjIzMA%3D%3D&v=1&f=sd', 'duration': 51, 'fallback_url': 'https://v.redd.it/99r6pskqem8g1/CMAF_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/99r6pskqem8g1/HLSPlaylist.m3u8?a=1768942553%2CMzMzNzFjZTc2NTgzNmExYTFiMDI3NzJiYzAwMzYyZjA5MjQ3NmNlNmQxYzIyZDNkNmMyN2NlNGJmNzdmODBmYQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/99r6pskqem8g1/CMAF_96.mp4', 'transcoding_status': 'completed', 'width': 1920}} | t3_1psgu8b | /r/LocalLLaMA/comments/1psgu8b/my_local_agent_built_this_stealth_game_in_one_go/ | false | false | 0 | {'enabled': False, 'images': [{'id': 'ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=108&crop=smart&format=pjpg&auto=webp&s=7718e046dd28b136c309fdc13ac93768076b5a2d', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=216&crop=smart&format=pjpg&auto=webp&s=564f8dacf9cfc012ec6146dba884eba5647d33c3', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=320&crop=smart&format=pjpg&auto=webp&s=dccd0927442d6fabbb409baab2ae11d77d062994', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=640&crop=smart&format=pjpg&auto=webp&s=e1f766029b28a82e755b8e490b344ca1e12ef885', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=960&crop=smart&format=pjpg&auto=webp&s=3071687d1de64a9c48b6f8aa03e2f849c6e7f48a', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?width=1080&crop=smart&format=pjpg&auto=webp&s=96d76e059cf2c0f39dcd16f7dc6f4f71b93c075d', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/ZnV1YmFrcHFlbThnMbAx8nw-haBMqpHU8uMF0BTU1OL2PHX3BGtYaHk_p9hR.png?format=pjpg&auto=webp&s=9c06b35f6268a008c48348e67ed2f87548f86711', 'width': 1920}, 'variants': {}}]} | |
Which tool should I pick to vibe code an app? | 0 | I’m looking for some advice from devs who actually use these tools day to day
I wanna vibe code a small app, nothing serious, mostly for fun and learning
The goal is to keep the flow smooth and not overthink everything
I’ve been checking out a few options so far:
Antrigravity
Claude
BlackBox
Windsurf
They all look solid in their own way, but it’s hard to understand the real tradeoffs without spending weeks on each one
If you had to pick one for vibe coding an app from scratch, which would you go with and why?
What worked well for you and what ended up being annoying?
Looking for real advice and honest experiences! Thanks in advance fam:) | 2025-12-21T20:50:40 | https://www.reddit.com/r/LocalLLaMA/comments/1psgq9s/which_tool_should_i_pick_to_vibe_code_an_app/ | viralgenius | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psgq9s | false | null | t3_1psgq9s | /r/LocalLLaMA/comments/1psgq9s/which_tool_should_i_pick_to_vibe_code_an_app/ | false | false | self | 0 | null |
I built a website that aggregates daily the latest challenges in GenAI research based on the latest published papers and list them as ideas anyone can work on them … I hope someone finds this useful. | 1 | [removed] | 2025-12-21T20:30:25 | PrincipleBest8031 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psg9ju | false | null | t3_1psg9ju | /r/LocalLLaMA/comments/1psg9ju/i_built_a_website_that_aggregates_daily_the/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': 'e6k62nhw9m8g1', 'resolutions': [{'height': 44, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=108&crop=smart&auto=webp&s=be9ee8d838a26b14349b165440ae6d1f6365d6a5', 'width': 108}, {'height': 88, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=216&crop=smart&auto=webp&s=7ae2e5e581111999c97a67ab40b4d91421036038', 'width': 216}, {'height': 131, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=320&crop=smart&auto=webp&s=79061c7759243694e0ea9c37f79b17ac41022989', 'width': 320}, {'height': 263, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=640&crop=smart&auto=webp&s=5ea0733a5616a69bc4821dd1edd88a704ba69817', 'width': 640}, {'height': 394, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=960&crop=smart&auto=webp&s=2ba15dbbb450492ac38605c0a3e492d32c177edf', 'width': 960}, {'height': 443, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?width=1080&crop=smart&auto=webp&s=6983cb671a1b7b9399f90d012117af392d41fc06', 'width': 1080}], 'source': {'height': 734, 'url': 'https://preview.redd.it/e6k62nhw9m8g1.png?auto=webp&s=15a6874c42580e6fd2e80edb9d6a1a8e9fe56d7f', 'width': 1786}, 'variants': {}}]} | |
AI REAL USEFUL WORKING IN REAL LIFE , LLAMA.CPP | 0 | hello i llama.cpp can work with excel files as , for example pdf , image etc , because excel files are the best form of working now for the office , there is no valid ia for real working if the ia cant handle and work with excel files!!!! there are a lot of free libs to work with excel files in python , how is the problem to develop this useful and fantastic idea to work in the office all the companyes and businnes men with using the ai to handle and make a lot of things as make the formulas very easy , format data , etc etc using this free python librarys and ai , llama.cpp must have a add excel file option as have to read pdf or jpg , txt , doc , etc excel is the most important file format for REAL WORKING IN THE REAL LIFE!!! | 2025-12-21T20:16:43 | https://www.reddit.com/r/LocalLLaMA/comments/1psfydn/ai_real_useful_working_in_real_life_llamacpp/ | Icy_Resolution8390 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psfydn | false | null | t3_1psfydn | /r/LocalLLaMA/comments/1psfydn/ai_real_useful_working_in_real_life_llamacpp/ | false | false | self | 0 | null |
Why so few open source multi modal llm, cost? | 2 | Was just wondering why so few multi modal llms that do image and voice/sound?
Is it cause of training cost?
Is it less of a market for it as most willing paying enterprises really just mostly need tool calling text?
Is it model size is too big for average user or enterprise to run?
Too complex?
When adding all 3 the modals, intelligence takes too big of a hit?
Don't get me wrong this has been a GREAT year for open source with many amazing models released and qwen released their 3 omni model which is all 3 modals. But it seems like only they released one. So I was curious what the main hurdle is.
Every few weeks I see poeple asking for a speaking model or how to do specs to text and text to speech. At least at hobby level seems their is interest.
| 2025-12-21T19:40:39 | https://www.reddit.com/r/LocalLLaMA/comments/1psf3i1/why_so_few_open_source_multi_modal_llm_cost/ | mindwip | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psf3i1 | false | null | t3_1psf3i1 | /r/LocalLLaMA/comments/1psf3i1/why_so_few_open_source_multi_modal_llm_cost/ | false | false | self | 2 | null |
ModelGuessr: Can you tell which AI you're chatting with? | 6 | I built ModelGuessr, a game where you chat with a random AI model (GPT 5.1, Sonnet 4.5, Gemini 2.5 Flash, Grok 4.1) and try to guess which one it is.
A big open question in AI is whether there's enough brand differentiation for AI companies to capture real profits. Will models end up commoditized like airline travel or differentiated like smartphones?
I built ModelGuessr to test this. I think that people will struggle more than they expect. And the more model mix-ups there are, the more commodity-like these models probably are.
If enough people play, I'll publish some follow-up analyses on confusion patterns (which models get mistaken for each other, what gives them away, etc.). Would love any feedback! | 2025-12-21T19:21:00 | https://model-guessr.com/ | kavalambda | model-guessr.com | 1970-01-01T00:00:00 | 0 | {} | 1psempi | false | null | t3_1psempi | /r/LocalLLaMA/comments/1psempi/modelguessr_can_you_tell_which_ai_youre_chatting/ | false | false | default | 6 | {'enabled': False, 'images': [{'id': 'CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4.png?width=108&crop=smart&auto=webp&s=0815170dacc12d1d0fdad271a4cbba69d24a9cec', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4.png?width=216&crop=smart&auto=webp&s=99ed329d5821b123de9e439e3a48776b9d5a5a94', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4.png?width=320&crop=smart&auto=webp&s=fc431efd8f917e4c10021d41eb8d707429b5e512', 'width': 320}, {'height': 335, 'url': 'https://external-preview.redd.it/CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4.png?width=640&crop=smart&auto=webp&s=e257da842477cf2e2f338acbaa36d31ced2f3a1a', 'width': 640}], 'source': {'height': 472, 'url': 'https://external-preview.redd.it/CszxA-KONv0l6iimAgbiztZ3sBEwzs6L_6XFOUQIRq4.png?auto=webp&s=2e0fa9f32f35805a6caa0d15fe2901cf408644a0', 'width': 900}, 'variants': {}}]} |
It ain’t much, but proud of my 2x3090 + a spare 3060 for support | 121 |
It’s a bit tight, but it fits and I didn’t want to buy a new case just yet. I had a spare computer that I bought first 1x3090, and now a 2nd 3090.
Qwen3-Next-80b is great!
Trying to wrap my head around Clint and using it in VS Code, but still not working properly… | 2025-12-21T19:03:54 | https://www.reddit.com/gallery/1pse7w6 | liviuberechet | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1pse7w6 | false | null | t3_1pse7w6 | /r/LocalLLaMA/comments/1pse7w6/it_aint_much_but_proud_of_my_2x3090_a_spare_3060/ | false | false | 121 | null | |
Open-source full-stack template for local LLM apps: FastAPI + Next.js, with LangChain/PydanticAI agents and multi-model support | 0 | Hey r/LocalLLaMA,
I've created an open-source project generator for building full-stack applications around LLMs – perfect for local setups, with support for running models like those from OpenAI/Anthropic (but easily extensible to local models via LangChain integrations). It's designed for rapid prototyping of chatbots, assistants, or ML tools with production infrastructure.
Repo: [https://github.com/vstorm-co/full-stack-fastapi-nextjs-llm-template](https://github.com/vstorm-co/full-stack-fastapi-nextjs-llm-template?referrer=grok.com)
(Install via pip install fastapi-fullstack, generate with fastapi-fullstack new – pick LangChain for broader LLM flexibility)
**LLM-focused features:**
* AI agents via LangChain (just added – with LangGraph for ReAct agents, tools, chains) or PydanticAI (type-safe with dependency injection)
* Multi-model support: Configure for local LLMs by swapping providers; streaming responses, conversation persistence, custom tools (e.g., database/external API access)
* Observability: LangSmith for LangChain traces (token usage, runs, feedback) or Logfire – great for debugging local model performance
* Backend: FastAPI for async handling, databases for history storage, background tasks for processing
* Frontend: Optional Next.js 15 chat UI with real-time WebSockets, dark mode, and tool visualizations
* DevOps: Docker for local deploys, Kubernetes manifests, and 20+ integrations (Redis, webhooks, etc.) to make local testing/production smooth
While it defaults to cloud models, the LangChain integration makes it easy to plug in local LLMs (e.g., via Ollama or HuggingFace). Screenshots (chat interfaces, LangSmith dashboards), demo GIFs, and AI docs in the README.
For local LLM devs:
* How does this fit with your setups for running models locally?
* Ideas for better local model support (e.g., specific integrations)?
* Pain points with full-stack LLM apps that this helps?
Contributions welcome – especially for local LLM enhancements! 🚀
Thanks! | 2025-12-21T18:50:47 | https://www.reddit.com/r/LocalLLaMA/comments/1psdwiw/opensource_fullstack_template_for_local_llm_apps/ | VanillaOk4593 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psdwiw | false | null | t3_1psdwiw | /r/LocalLLaMA/comments/1psdwiw/opensource_fullstack_template_for_local_llm_apps/ | false | false | self | 0 | null |
Nemotron-Nano-30B: What settings are you getting good results with? | 30 | Currently I'm running with the settings from the model card for tool-calling:
- temperature=0.6
- top_p=0.95
- top_k 20
Everything goes well until you're about 50k tokens in, then it kind of goes off the rails, enters infinite retry loops, or starts doing things that I can only describe as *"silly"*.
My use-case is agentic coding with Qwen-Code-CLI. | 2025-12-21T18:36:33 | https://www.reddit.com/r/LocalLLaMA/comments/1psdk9w/nemotronnano30b_what_settings_are_you_getting/ | EmPips | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psdk9w | false | null | t3_1psdk9w | /r/LocalLLaMA/comments/1psdk9w/nemotronnano30b_what_settings_are_you_getting/ | false | false | self | 30 | null |
Built a free local voice dictation app using faster-whisper instead of paying for Wispr Flow | 21 | I was using Wispr Flow for voice dictation and it worked well but then I thought wait, I have a GPU, I can just run Whisper locally instead of paying $12/month and sending my audio to their servers.
The app uses the CTranslate2 optimized Whisper models from Systran on Hugging Face. You can pick from tiny, base, small, medium, large-v3, or the distil-large-v3-turbo depending on your speed and accuracy needs. I've been running tiny and base on CPU and they're basically instant. Just added GPU/CUDA support in the latest version and the larger models are now near-instant too, honestly faster than cloud solutions in my testing.
Accuracy has been solid for my use case. Whisper benchmarks around 2-5% word error rate on clean English and I haven't had issues with it in daily use. No AI post-processing so you get raw transcription which is the main thing missing compared to Wispr Flow.
The app sits in your system tray, you hold a hotkey to record, release and it transcribes and pastes wherever your cursor is. Works system-wide so you can use it with any text field. Custom hotkeys and a toggle mode if you don't want to hold. Unlimited recording length. All data stays on your machine, nothing leaves your PC.
Here's how it works: [https://i.redd.it/zyy8aujqpk8g1.gif](https://i.redd.it/zyy8aujqpk8g1.gif)
I want to add text cleanup like Wispr Flow does but keep it runnable on CPU for people without GPUs. I've been looking at small models like Qwen2-0.5B, SmolLM-360M, Gemma 3 1B, and Phi-3 Mini. From what I've read Qwen2-0.5B and SmolLM can run unquantized on CPU with a small memory footprint, and with Q4 quantization even Qwen 1.5B or Gemma 3 1B might be viable. Has anyone done something similar for text post-processing? Can any of these realistically fit in under 300-500MB of RAM and still produce decent cleanup, or is GPU basically required for anything useful? Any recommendations?
Windows only for now. Free and open source under MIT.
Download: [https://get-voice-flow.vercel.app/](https://get-voice-flow.vercel.app/)
Source: [https://github.com/infiniV/VoiceFlow](https://github.com/infiniV/VoiceFlow) | 2025-12-21T18:36:09 | https://www.reddit.com/gallery/1psdjxc | raww2222 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1psdjxc | false | null | t3_1psdjxc | /r/LocalLLaMA/comments/1psdjxc/built_a_free_local_voice_dictation_app_using/ | false | false | 21 | null | |
As 2025 wraps up, which local LLMs really mattered this year and what do you want to see in 2026? | 63 | Now that we’re at the end of 2025, I’m curious how people here would summarize the local LLM landscape this year.
Not just “what scores highest on benchmarks,” but:
\- What models did people actually run?
\- What felt popular or influential in practice?
\- What models punched above their weight?
\- What disappointed or faded out?
Looking back, which local LLMs defined 2025 for you?
And looking forward:
\- What gaps still exist?
\- What do you want to see next year? (better small models, longer context, better reasoning, multimodal, agents, efficiency, etc.)
Would love both personal takes and broader ecosystem observations. | 2025-12-21T18:23:33 | https://www.reddit.com/r/LocalLLaMA/comments/1psd918/as_2025_wraps_up_which_local_llms_really_mattered/ | ObjectiveOctopus2 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psd918 | false | null | t3_1psd918 | /r/LocalLLaMA/comments/1psd918/as_2025_wraps_up_which_local_llms_really_mattered/ | false | false | self | 63 | null |
A list of 28 modern benchmarks and their short description | 27 | I realised that my understanding of the benchmarks was stuck somewhere around GSM8k/SimpleQA area - very dated by now.
So I went through some of the recent releases and compiled a list of the used benchmarks and what they represent. Some of these are very obvious (ARC-AGI, AIME, etc.) but for many - I was seeing them for the first time, so I hope it'll be useful for someone else too.
|Benchmark|Description|
|:-|:-|
|[**AIME 2025**](https://artificialanalysis.ai/evaluations/aime-2025)|Tests olympiad-level mathematical reasoning using all 30 problems from the 2025 American Invitational Mathematics Examination with integer answers from 000-999|
|[**ARC-AGI-1 (Verified)**](https://arcprize.org/arc-agi/1/)|Measures basic fluid intelligence through visual reasoning puzzles that are easy for humans but challenging for AI systems|
|[**ARC-AGI-2**](https://arcprize.org/arc-agi/2/)|An updated benchmark designed to stress test the efficiency and capability of state-of-the-art AI reasoning systems with visual pattern recognition tasks|
|[**CharXiv Reasoning**](https://charxiv.github.io/)|Evaluates information synthesis from complex charts through descriptive and reasoning questions that require analyzing visual elements|
|[**Codeforces**](https://arxiv.org/html/2501.01257v1)|A competition-level coding benchmark that evaluates LLM programming capabilities using problems from the CodeForces platform with standardized ELO ratings|
|[**FACTS Benchmark Suite**](https://deepmind.google/blog/facts-benchmark-suite-systematically-evaluating-the-factuality-of-large-language-models/)|Systematically evaluates Large Language Model factuality across parametric, search, and multimodal reasoning domains|
|[**FrontierMath (Tier 1-3)**](https://epoch.ai/frontiermath)|Tests undergraduate through early graduate level mathematics problems that take specialists hours to days to solve|
|[**FrontierMath (Tier 4)**](https://epoch.ai/frontiermath)|Evaluates research-level mathematics capabilities with exceptionally challenging problems across major branches of modern mathematics|
|[**GDPval**](https://openai.com/index/gdpval/)|Measures AI model performance on real-world economically valuable tasks across 44 occupations from the top 9 industries contributing to U.S. GDP|
|[**Global PIQA**](https://arxiv.org/abs/2510.24081)|Evaluates physical commonsense reasoning across over 100 languages with culturally-specific examples created by native speakers|
|[**GPQA Diamond**](https://arxiv.org/abs/2311.12022)|Tests graduate-level scientific knowledge through multiple-choice questions that domain experts can answer but non-experts typically cannot|
|[**HMMT 2025**](https://llm-stats.com/benchmarks/hmmt-2025)|Assesses mathematical reasoning using problems from the Harvard-MIT Mathematics Tournament, a prestigious high school mathematics competition|
|[**Humanity's Last Exam**](https://agi.safe.ai/)|A multi-modal benchmark designed to test expert-level performance on closed-ended, verifiable questions across dozens of academic subjects|
|[**LiveCodeBench Pro**](https://livecodebenchpro.com/)|Evaluates LLM code generation capabilities on competitive programming problems of varying difficulty levels from different platforms|
|[**MCP Atlas**](https://scale.com/leaderboard/mcp_atlas)|Measures how well language models handle real-world tool use through multi-step workflows using the Model Context Protocol|
|[**MMMLU**](https://huggingface.co/datasets/openai/MMMLU/blob/main/README.md)|A multilingual version of MMLU featuring professionally translated questions across 14 languages to test massive multitask language understanding|
|[**MMMU-Pro**](https://arxiv.org/abs/2409.02813)|A more robust multimodal benchmark that filters text-only answerable questions and augments options to test true multimodal understanding|
|[**MRCH v2 (8-needle)**](https://llm-stats.com/benchmarks/mrcr-v2-(8-needle))|Tests models' ability to simultaneously track and reason about 8 pieces of information across extended conversations in long contexts|
|[**OmniDocBench 1.5**](https://github.com/opendatalab/OmniDocBench)|Evaluates diverse document parsing capabilities across 9 document types, 4 layout types, and 3 languages with rich OCR annotations|
|[**ScreenSpot-Pro**](https://huggingface.co/datasets/Voxel51/ScreenSpot-Pro)|Assesses GUI grounding capabilities in high-resolution professional software environments across 23 applications and 5 industries|
|[**SimpleQA Verified**](https://arxiv.org/abs/2509.07968)|A reliable factuality benchmark with 1,000 prompts for evaluating short-form factual accuracy in Large Language Models|
|[**SWE-bench Pro (public)**](https://scale.com/leaderboard/swe_bench_pro_public)|A rigorous software engineering benchmark designed to address data contamination with more diverse and difficult coding tasks|
|[**SWE-bench Verified**](https://scale.com/blog/swe-bench-pro)|Tests agentic coding capabilities on verified software engineering problems with solutions that have been manually validated|
|[**t²-Bench**](https://artificialanalysis.ai/evaluations/tau2-bench)|A dual-control conversational AI benchmark simulating technical support scenarios where both agent and user coordinate actions|
|[**Terminal-bench 2.0**](https://www.tbench.ai/)|Measures AI agent capabilities in terminal environments through complex tasks like compiling code, training classifiers, and server setup|
|[**Toolathlon**](https://github.com/hkust-nlp/Toolathlon)|Benchmarks language agents' general tool use in realistic environments featuring 600+ diverse tools and long-horizon task execution|
|[**Vending-Bench 2**](https://andonlabs.com/evals/vending-bench-2)|Evaluates AI model performance on running a simulated vending machine business over long time horizons, scored on final bank balance|
|[**Video-MMMU**](https://videommmu.github.io/)|Assesses Large Multimodal Models' ability to acquire and utilize knowledge from expert-level videos across six disciplines| | 2025-12-21T18:20:13 | https://www.reddit.com/r/LocalLLaMA/comments/1psd61v/a_list_of_28_modern_benchmarks_and_their_short/ | Everlier | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psd61v | false | null | t3_1psd61v | /r/LocalLLaMA/comments/1psd61v/a_list_of_28_modern_benchmarks_and_their_short/ | false | false | self | 27 | null |
Moore Threads Unveils The Lushan Gaming & Huashan AI GPUs: 15x Gaming Performance Uplift, 50x RT Boost, DX12 Ultimate Support, Launching Next Year | 96 | https://wccftech.com/moore-threads-lushan-gaming-huashan-ai-gpus-15x-gaming-uplift-50x-rt-boost-dx12-ultimate-support/
| 2025-12-21T18:18:24 | https://www.reddit.com/r/LocalLLaMA/comments/1psd4er/moore_threads_unveils_the_lushan_gaming_huashan/ | Individual_Aside7554 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psd4er | false | null | t3_1psd4er | /r/LocalLLaMA/comments/1psd4er/moore_threads_unveils_the_lushan_gaming_huashan/ | false | false | self | 96 | {'enabled': False, 'images': [{'id': '5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=108&crop=smart&auto=webp&s=74420c3278e44f076532bf8ebff1b301cfc95ec1', 'width': 108}, {'height': 145, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=216&crop=smart&auto=webp&s=4e00ef8ca735d9ab6d3e28628a0a2c23364835af', 'width': 216}, {'height': 216, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=320&crop=smart&auto=webp&s=abd441bf8fbf25b3989aa956e004724152eb3966', 'width': 320}, {'height': 432, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=640&crop=smart&auto=webp&s=1f251edf5aa0a206d795ec824fe35129880ef8b3', 'width': 640}, {'height': 648, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=960&crop=smart&auto=webp&s=99ac188a92b70ed6a7330308e1cf9205634e2f0d', 'width': 960}, {'height': 729, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?width=1080&crop=smart&auto=webp&s=a8c399bfb50bc179afde43eaa2b37e91f0692f6d', 'width': 1080}], 'source': {'height': 1304, 'url': 'https://external-preview.redd.it/5duqin65q-ozx9siEuRgFdMk7Exo5Vl4didLtZQDOHA.png?auto=webp&s=4037846bfcfa4331562d3b56c062c3b84110f8cb', 'width': 1930}, 'variants': {}}]} |
Local LLMs on potato computers feat. the llm Python CLI and sllm.nvim, and why you should stop using big bloated AI tools | 4 | Hello LocalLLaMA!
I've been following the sub for years at this point but never really ran any LLM myself. Most models are just too big: I simply can't run them on my laptop. But these last few weeks, I've been trying out a local setup using Ollama, the llm Python CLI and the sllm.nvim plugin, small models, and have been pretty impressed at what they can do. Small LLMs are getting insanely good.
I share my setup and various tips and tricks in this article:
[https://zoug.fr/local-llms-potato-computers/](https://zoug.fr/local-llms-potato-computers/)
It's split into two parts. A first one, technical, where I share my setup (the one linked above) but also a second, non-technical one where I talk about the AI bubble, the environmental costs of LLMs and the true benefits of using AI as a programmer/computer engineer:
[https://zoug.fr/stop-using-big-bloated-ai/](https://zoug.fr/stop-using-big-bloated-ai/)
I'm very interested in your feedback. I know what I'm saying in these articles is probably not what most people here think, so all the more reason. I hope you'll get something out of them! Thanks :) | 2025-12-21T18:15:49 | https://www.reddit.com/r/LocalLLaMA/comments/1psd24y/local_llms_on_potato_computers_feat_the_llm/ | yzoug | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psd24y | false | null | t3_1psd24y | /r/LocalLLaMA/comments/1psd24y/local_llms_on_potato_computers_feat_the_llm/ | false | false | self | 4 | {'enabled': False, 'images': [{'id': 'FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=108&crop=smart&auto=webp&s=545016bb209cf7776cd5ab4b75ac521d9d665c22', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=216&crop=smart&auto=webp&s=7c905c487b43da6db05c96293a02f49d266e376c', 'width': 216}, {'height': 167, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=320&crop=smart&auto=webp&s=ae4da2bbfed869e989e956e2851242a88682ff55', 'width': 320}, {'height': 334, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=640&crop=smart&auto=webp&s=a79c04a94b0f013ee93599993d315c9c74fe8479', 'width': 640}, {'height': 502, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=960&crop=smart&auto=webp&s=639b043edfb50b8a36338d46af2fe77b7d356388', 'width': 960}, {'height': 565, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?width=1080&crop=smart&auto=webp&s=9c45c12e3ad39e6f55c50338f3450eb0fa3f01c3', 'width': 1080}], 'source': {'height': 628, 'url': 'https://external-preview.redd.it/FFIPuNDwNPcBZP7cwoulke1tXsfizdIfZwia8oGseUQ.jpeg?auto=webp&s=a11020d7938a521ad3f4251435aff1ee69a20434', 'width': 1200}, 'variants': {}}]} |
How does the models in Openrouter work? | 0 | So far I've been using Openrouter for roleplay and its enjoyable. So far like Grok 4.1, when the credits are insufficient to continue with them, is it like fully over or they refill? And what model is good for manga/canon accurate roleplays with the theme and its tone?
Correct me if im wrong. | 2025-12-21T18:14:41 | https://www.reddit.com/r/LocalLLaMA/comments/1psd15s/how_does_the_models_in_openrouter_work/ | MajimaLovesKiryu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psd15s | false | null | t3_1psd15s | /r/LocalLLaMA/comments/1psd15s/how_does_the_models_in_openrouter_work/ | false | false | self | 0 | null |
Fck OpenAI honestly | 47 | 2025-12-21T17:56:29 | Slight_Tone_2188 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1pscl8i | false | null | t3_1pscl8i | /r/LocalLLaMA/comments/1pscl8i/fck_openai_honestly/ | false | false | default | 47 | {'enabled': True, 'images': [{'id': 'ti8skg12jl8g1', 'resolutions': [{'height': 71, 'url': 'https://preview.redd.it/ti8skg12jl8g1.jpeg?width=108&crop=smart&auto=webp&s=ab7d768d47bc328c614022b91bbd2f915332e55f', 'width': 108}, {'height': 143, 'url': 'https://preview.redd.it/ti8skg12jl8g1.jpeg?width=216&crop=smart&auto=webp&s=42078ab4cc3af486218a7559004ea8a820c56791', 'width': 216}, {'height': 213, 'url': 'https://preview.redd.it/ti8skg12jl8g1.jpeg?width=320&crop=smart&auto=webp&s=fe386f6246cd1beb2b26ad6b4b0b0be2d7ac5739', 'width': 320}, {'height': 426, 'url': 'https://preview.redd.it/ti8skg12jl8g1.jpeg?width=640&crop=smart&auto=webp&s=439553f18cb306d3cfd2237869dd0e7a77953886', 'width': 640}], 'source': {'height': 442, 'url': 'https://preview.redd.it/ti8skg12jl8g1.jpeg?auto=webp&s=78a5734255c1c4642298cb3c5d25f4e7e7cfb12e', 'width': 664}, 'variants': {}}]} | ||
[Day 1] Can hobbyists design a robust LLM? Making a community open-source LLM, top comments get implemented. | 1 | [removed] | 2025-12-21T17:50:43 | https://www.reddit.com/r/LocalLLaMA/comments/1pscgbj/day_1_can_hobbyists_design_a_robust_llm_making_a/ | Shizuka_Kuze | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1pscgbj | false | null | t3_1pscgbj | /r/LocalLLaMA/comments/1pscgbj/day_1_can_hobbyists_design_a_robust_llm_making_a/ | false | false | self | 1 | null |
Best hardware now or in Q1 26 to run local LLMs for text analysis? | 2 | Hi everyone,
I'm trying to get an overview of hardware options but I'm very new to local LLMs and frankly overwhelmed by all the choices. Would really appreciate some guidance from folks who've been through this.
I've been running 7-8B models on my M1 MacBook (16GB) through LMStudio. Works fine for rewriting emails but it's useless for what I actually need - analysing very long and many interview transcripts and doing proper text based research. I tried running bigger models on a HPC cluster but honestly the whole SSH'ing, job queue, waiting around thing just kills my workflow. I would like to iterate quickly, run agents, pass data between processing steps. And all that locally, accessible via phone / laptop would be the dream.
I'm doing heavy text analysis work from March until September 2026 so i was thinking of just buying my own hardware. Budget available is around 2-3k euro. I travel every few months so those small desktop AI PCs caught my eye - the DGX Spark or its siblings, Framework or other AI 365 mashines, Mac Mini M4 Pro, maybe Mac Studio. Not sure which platform would work best for remoting in from my macbook or using openweb ui. Regarding the mini I keep asking myself will 48 or 64GB be enough or will i immediately wish i had more? The 128GB unified ram option can run the 200B models, which would be neat, but I don't know if another platform (linux? windows?) is going to be a pain.
Adding to my confusion: i see people here casually talking about their Mac Studios with 256 or 512GB like that's normal, which makes 48GB sound pathetic. Those are 6k+ which i can't afford right now but could save up for by mid-2026. And then there's the M5 Max/Ultra possibly coming Q3 2026. So is it smarter to buy something 'cheap' now for 2k to learn and experiment, then upgrade to a beast later? Or will that just be wasting money on two systems? Also not sure how much RAM i actually need for my use case. I want to run really nuanced models for analyzing transcripts, maybe some agent workflows with different 'analyst roles'. What amount of RAM do I really need? Anyone doing similar work who can share what actually works in practice?
thanks from a lost soul :D | 2025-12-21T17:35:14 | https://www.reddit.com/r/LocalLLaMA/comments/1psc337/best_hardware_now_or_in_q1_26_to_run_local_llms/ | DjFlu | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psc337 | false | null | t3_1psc337 | /r/LocalLLaMA/comments/1psc337/best_hardware_now_or_in_q1_26_to_run_local_llms/ | false | false | self | 2 | null |
llama.cpp appreciation post | 1,529 | 2025-12-21T17:28:24 | hackiv | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1psbx2q | false | null | t3_1psbx2q | /r/LocalLLaMA/comments/1psbx2q/llamacpp_appreciation_post/ | false | false | default | 1,529 | {'enabled': True, 'images': [{'id': 'asipaua1el8g1', 'resolutions': [{'height': 132, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=108&crop=smart&auto=webp&s=36d4321d8c882699567fade589cf90d90e4417f3', 'width': 108}, {'height': 265, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=216&crop=smart&auto=webp&s=65497356708104c35f8b0e198d62582d1539ba9a', 'width': 216}, {'height': 393, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=320&crop=smart&auto=webp&s=11d75c1ed75678c8be8e5a932f8d563adab8bae9', 'width': 320}, {'height': 786, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=640&crop=smart&auto=webp&s=87eeb0e85f39e765b810e9ec58e5148346cc419b', 'width': 640}, {'height': 1179, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=960&crop=smart&auto=webp&s=04da6ab711816abaeb9126f86f87b7c3132f9a12', 'width': 960}, {'height': 1327, 'url': 'https://preview.redd.it/asipaua1el8g1.png?width=1080&crop=smart&auto=webp&s=6e52adaac27df20b20b898f44c8437ed94d71a02', 'width': 1080}], 'source': {'height': 1327, 'url': 'https://preview.redd.it/asipaua1el8g1.png?auto=webp&s=dbffce99c3eefd5feefac944cd60e3839403a11c', 'width': 1080}, 'variants': {}}]} | ||
Suggestions for open-source models for English->Hindi translation | 0 | I was trying to translate one of the dialogues from Deadpool
"Mr. Wilson, you appear to have soiled yourself while on duty."
and without promping, the best translation that I could get was Grok (yeah I know its closed-source but I wanted to try)
"मिस्टर विल्सन, लगता है कि आपने ड्यूटी के दौरान खुद को गंदा कर लिया है।"
But this totally changes the tone/intent of the dialogue. I could only get better results when I prompted Gemini3.0-Fast, giving it context of the movie. The results are much better and acceptable:
\`\`\`
**"मिस्टर विल्सन, लगता है ड्यूटी पर रहते हुए आपने अपनी पैंट गीली कर दी है।"**
अगर आप इसे थोड़ा और डेडपूल वाले मज़ाकिया या 'R-rated' लहजे में चाहते हैं, तो यह भी कह सकते हैं:
**"मिस्टर विल्सन, ऐसा लग रहा है कि ड्यूटी के दौरान आपका 'काम' आपकी पैंट में ही तमाम हो गया है।"**
**\`\`\`**
I still need a local/open-source model that requires minimal prompting. | 2025-12-21T17:13:37 | https://www.reddit.com/r/LocalLLaMA/comments/1psbk6g/suggestions_for_opensource_models_for/ | Warm-Professor-9299 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1psbk6g | false | null | t3_1psbk6g | /r/LocalLLaMA/comments/1psbk6g/suggestions_for_opensource_models_for/ | false | false | nsfw | 0 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.