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 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Just finished building this bad boy | 241 | 6x Gigabyte 3090 Gaming OC all running at PCIe 4.0 16x speed
Asrock Romed-2T motherboard with Epyc 7502 CPU
8 sticks of DDR4 8GB 2400Mhz running in octochannel mode
Modified Tinygrad Nvidia drivers with P2P enabled, intra GPU bandwidth tested at 24.5 GB/s
Total 144GB VRam, will be used to experiment with training d... | 2026-02-11T10:28:00 | dazzou5ouh | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1tuh1 | false | null | t3_1r1tuh1 | /r/LocalLLaMA/comments/1r1tuh1/just_finished_building_this_bad_boy/ | false | false | 241 | {'enabled': True, 'images': [{'id': 'ju0ed5uceuig1', 'resolutions': [{'height': 80, 'url': 'https://preview.redd.it/ju0ed5uceuig1.jpeg?width=108&crop=smart&auto=webp&s=318cfc5daf59aca2e1d3b4859805aa8a3d1e343c', 'width': 108}, {'height': 161, 'url': 'https://preview.redd.it/ju0ed5uceuig1.jpeg?width=216&crop=smart&auto=w... | ||
Any local 70B model or less that comes close to gemini flash lite? | 1 | As of today, I mean
I still haven't seen anything that comes close to gemini for text summarization. Locally at least | 2026-02-11T10:21:19 | https://www.reddit.com/r/LocalLLaMA/comments/1r1tqia/any_local_70b_model_or_less_that_comes_close_to/ | goingsplit | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1tqia | false | null | t3_1r1tqia | /r/LocalLLaMA/comments/1r1tqia/any_local_70b_model_or_less_that_comes_close_to/ | false | false | self | 1 | null |
Behavioral probe on epistemic responsibility in 4 LLMs + open standard proposal (Anchor v0.1) | 0 | I’ve been running a small behavior-focused probe to test how current LLMs handle epistemic stress situations that require uncertainty disclosure, bounded recall, or reframing invalid premises.
The goal wasn’t to rank models or estimate prevalence.
The goal was to identify repeatable failure classes under specific pr... | 2026-02-11T10:08:09 | https://www.reddit.com/r/LocalLLaMA/comments/1r1tigm/behavioral_probe_on_epistemic_responsibility_in_4/ | Lost-Albatross5241 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1tigm | false | null | t3_1r1tigm | /r/LocalLLaMA/comments/1r1tigm/behavioral_probe_on_epistemic_responsibility_in_4/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'ewqWNe7uw0Z2ayLmqIjCyK89O-K938RQhf0e_V2VPMI', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ewqWNe7uw0Z2ayLmqIjCyK89O-K938RQhf0e_V2VPMI.png?width=108&crop=smart&auto=webp&s=0e3596985b37031db933bd66052db6717f6bcee2', 'width': 108}, {'height': 108, 'url': 'h... |
[Project Share] I built a Python "OS" to orchestrate ComfyUI & Ollama locally. No more gacha, just pipelines. (Open Source) | 1 | **Hi everyone,**
I'm a liberal arts graduate who taught myself Python because I was frustrated with the inconsistency of AI video generation.
I built **DarkFactory**, an automation framework that treats ComfyUI models as "cartridges" and uses Python + Ollama as the "Director".
**Repo:**[https://github.com/maikarlone... | 2026-02-11T10:05:04 | https://www.reddit.com/r/LocalLLaMA/comments/1r1tgl0/project_share_i_built_a_python_os_to_orchestrate/ | Popular_Teacher4231 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1tgl0 | false | null | t3_1r1tgl0 | /r/LocalLLaMA/comments/1r1tgl0/project_share_i_built_a_python_os_to_orchestrate/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'fC8lieOPtkcEAtfLV8rn1EDvDW0H_oDoBMSM14-DU3I', 'resolutions': [{'height': 47, 'url': 'https://external-preview.redd.it/fC8lieOPtkcEAtfLV8rn1EDvDW0H_oDoBMSM14-DU3I.jpeg?width=108&crop=smart&auto=webp&s=2cf6b06a63adccfcb45482f378a542e157181ab0', 'width': 108}, {'height': 94, 'url': 'h... |
DeepSeek just updated to a 1M context window! | 47 | The DeepSeek app was just updated with 1M context, and the knowledge cutoff date is now May 2025. It's unclear for now if this is a new model. Also, there hasn't been any movement on their Hugging Face page yet.
https://preview.redd.it/9z2ggdgy9uig1.png?width=1179&format=png&auto=webp&s=a3f48da856b53751f2db2b17ac5f49b... | 2026-02-11T10:03:06 | https://www.reddit.com/r/LocalLLaMA/comments/1r1tfbm/deepseek_just_updated_to_a_1m_context_window/ | Dr_Karminski | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1tfbm | false | null | t3_1r1tfbm | /r/LocalLLaMA/comments/1r1tfbm/deepseek_just_updated_to_a_1m_context_window/ | false | false | 47 | null | |
Pony Alpha Uncloaked. | 0 | I
https://preview.redd.it/j1fpkzln7uig1.png?width=975&format=png&auto=webp&s=0446b9496dcc0515ed77e278420f87f3be4a71f6
Comparable quality.... Superior quality is what it meant. | 2026-02-11T09:50:59 | https://www.reddit.com/r/LocalLLaMA/comments/1r1t7u4/pony_alpha_uncloaked/ | volious-ka | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1t7u4 | false | null | t3_1r1t7u4 | /r/LocalLLaMA/comments/1r1t7u4/pony_alpha_uncloaked/ | false | false | 0 | null | |
I built an MCP server that gives AI agents full control of Windows desktops (40+ tools, open source) | 5 | I got frustrated with the lack of proper Windows support in the MCP ecosystem, so I built WinRemote MCP — an open-source MCP server that lets AI agents control Windows machines remotely.
What it does:
• Screenshots with UI element detection + OCR
• Mouse/keyboard control (click, type, scroll, shortcuts)
• File ... | 2026-02-11T09:37:58 | https://www.reddit.com/r/LocalLLaMA/comments/1r1t0dt/i_built_an_mcp_server_that_gives_ai_agents_full/ | Neat-Football1149 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1t0dt | false | null | t3_1r1t0dt | /r/LocalLLaMA/comments/1r1t0dt/i_built_an_mcp_server_that_gives_ai_agents_full/ | false | false | self | 5 | {'enabled': False, 'images': [{'id': 'qoNHoVL3kmex1BTtLa34lxIzTIk8ZmhOo2MVa10e9w8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/qoNHoVL3kmex1BTtLa34lxIzTIk8ZmhOo2MVa10e9w8.png?width=108&crop=smart&auto=webp&s=999244082b6173bcb70bfe31ae4b3e9169283dbe', 'width': 108}, {'height': 108, 'url': 'h... |
I built an autonomous research agent in C# that runs entirely on local LLMs (Ollama + llama3.1:8b) | 0 | I got tired of manually copy-pasting URLs into ChatGPT for research, so I built an agent that does it autonomously. Figured I'd share since this sub loves practical local LLM projects.
**What it does:**
- You give it a topic ("persistent memory for AI agents")
- It generates 5-8 search queries
- Searches the web via B... | 2026-02-11T09:16:44 | https://www.reddit.com/r/LocalLLaMA/comments/1r1soca/i_built_an_autonomous_research_agent_in_c_that/ | Dynamic-Styles | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1soca | false | null | t3_1r1soca | /r/LocalLLaMA/comments/1r1soca/i_built_an_autonomous_research_agent_in_c_that/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': '8JUT6t06UR8i9-X8wMLqZOpBlnveOJ0qlQ_wGC38fjg', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/8JUT6t06UR8i9-X8wMLqZOpBlnveOJ0qlQ_wGC38fjg.png?width=108&crop=smart&auto=webp&s=fb3508b60a4a77c4a8d02e85856fb37551d9bc07', 'width': 108}, {'height': 108, 'url': 'h... |
DeepSeek has launched grayscale testing for its new model on both its official website and app. 1M content length! | 127 | DeepSeek has launched grayscale testing for its new model on both its official website and app. The new model features a 1M context window and an updated knowledge base. Currently, access is limited to a select group of accounts."
https://preview.redd.it/j1qiarng1uig1.png?width=1163&format=png&auto=webp&s=3a99f1652ea7... | 2026-02-11T09:15:17 | https://www.reddit.com/r/LocalLLaMA/comments/1r1snhv/deepseek_has_launched_grayscale_testing_for_its/ | External_Mood4719 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1snhv | false | null | t3_1r1snhv | /r/LocalLLaMA/comments/1r1snhv/deepseek_has_launched_grayscale_testing_for_its/ | false | false | 127 | null | |
Step-3.5-Flash AIME 2026 Results | 48 | 2026-02-11T09:14:14 | https://www.reddit.com/r/LocalLLaMA/comments/1r1smw0/step35flash_aime_2026_results/ | Abject-Ranger4363 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1smw0 | false | null | t3_1r1smw0 | /r/LocalLLaMA/comments/1r1smw0/step35flash_aime_2026_results/ | false | false | 48 | null | ||
I am planning on building a home AI server, what would you recommend | 1 | I have seen many build around this price before ram surge, my budget is around 2500 USD not counting ram. I will try and read all your recommendations! | 2026-02-11T08:58:28 | https://www.reddit.com/r/LocalLLaMA/comments/1r1sdkp/i_am_planning_on_building_a_home_ai_server_what/ | RecognitionPatient12 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1sdkp | false | null | t3_1r1sdkp | /r/LocalLLaMA/comments/1r1sdkp/i_am_planning_on_building_a_home_ai_server_what/ | false | false | self | 1 | null |
I have 24GB VRAM and 64-72GB system memory. What coding model for a newbie would you recommend? | 6 | Title. A buddy of mine is running rnj-1 8b. I always read that qwen coder 3 was pretty top tier. Just read some posts that said it wasn't that great and running into issues. I don't have any projects in mind but somewhere between batch and bash scripting I think I could learn some more. Preferably python. Thanks in adv... | 2026-02-11T08:24:16 | https://www.reddit.com/r/LocalLLaMA/comments/1r1rtt3/i_have_24gb_vram_and_6472gb_system_memory_what/ | ziggo0 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1rtt3 | false | null | t3_1r1rtt3 | /r/LocalLLaMA/comments/1r1rtt3/i_have_24gb_vram_and_6472gb_system_memory_what/ | false | false | self | 6 | null |
Implemented Pipeline Parallelism w/ micro batching on 3x Mac minis! | 1 | [removed] | 2026-02-11T07:47:22 | https://www.reddit.com/r/LocalLLaMA/comments/1r1r8ik/implemented_pipeline_parallelism_w_micro_batching/ | East-Muffin-6472 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1r8ik | false | null | t3_1r1r8ik | /r/LocalLLaMA/comments/1r1r8ik/implemented_pipeline_parallelism_w_micro_batching/ | false | false | 1 | null | |
Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts | 144 | Hi everyone 👋
We’re excited to share Nanbeige4.1-3B, the latest iteration of our open-source 3B model from Nanbeige LLM Lab. Our goal with this release is to explore whether a small general model can simultaneously achieve strong reasoning, robust preference alignment, and agentic behavior.
https://preview.redd.it/8... | 2026-02-11T07:38:59 | https://www.reddit.com/r/LocalLLaMA/comments/1r1r3nk/nanbeige413b_a_small_general_model_that_reasons/ | Tiny_Minimum_4384 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1r3nk | false | null | t3_1r1r3nk | /r/LocalLLaMA/comments/1r1r3nk/nanbeige413b_a_small_general_model_that_reasons/ | false | false | 144 | null | |
Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts | 2 | Hi everyone 👋
We’re excited to share Nanbeige4.1-3B, the latest iteration of our open-source 3B model from Nanbeige LLM Lab. Our goal with this release is to explore whether a small general model can simultaneously achieve strong reasoning, robust preference alignment, and agentic behavior.
[https://preview.redd.it/... | 2026-02-11T07:35:23 | https://www.reddit.com/r/LocalLLaMA/comments/1r1r1kn/nanbeige413b_a_small_general_model_that_reasons/ | Tiny_Minimum_4384 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1r1kn | false | null | t3_1r1r1kn | /r/LocalLLaMA/comments/1r1r1kn/nanbeige413b_a_small_general_model_that_reasons/ | false | false | self | 2 | {'enabled': False, 'images': [{'id': 'YjyrWqfDJVwgGy5czS9cV1hWN9BL7-RLvY00hNEmG0M', 'resolutions': [{'height': 73, 'url': 'https://external-preview.redd.it/YjyrWqfDJVwgGy5czS9cV1hWN9BL7-RLvY00hNEmG0M.png?width=108&crop=smart&auto=webp&s=14ab9283b35c5234e5427a4ebb3fe398524db042', 'width': 108}, {'height': 146, 'url': 'h... |
Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts | 1 | [removed] | 2026-02-11T07:31:31 | https://www.reddit.com/r/LocalLLaMA/comments/1r1qzdg/nanbeige413b_a_small_general_model_that_reasons/ | leran2098 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1qzdg | false | null | t3_1r1qzdg | /r/LocalLLaMA/comments/1r1qzdg/nanbeige413b_a_small_general_model_that_reasons/ | false | false | self | 1 | null |
GLM-4.7.Flash - is it normal to behave like that? It's like I am talking to my anxious, Chinese girlfriend. I don't use AI so this is new to me | 8 | 2026-02-11T07:27:49 | Mayion | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1qx4j | false | null | t3_1r1qx4j | /r/LocalLLaMA/comments/1r1qx4j/glm47flash_is_it_normal_to_behave_like_that_its/ | false | false | default | 8 | {'enabled': True, 'images': [{'id': 'orx5tik1itig1', 'resolutions': [{'height': 77, 'url': 'https://preview.redd.it/orx5tik1itig1.jpeg?width=108&crop=smart&auto=webp&s=619813a641d12eebe7c8f09316babdc8aaa8465b', 'width': 108}, {'height': 154, 'url': 'https://preview.redd.it/orx5tik1itig1.jpeg?width=216&crop=smart&auto=w... | ||
Dual RTX 5060 Ti (32GB pooled VRAM) vs Single RTX 5070 Ti (16GB): Real-world LLM benchmarks on Blackwell | 4 | I am the obsessive sort, and lately my obsession is ML/AI and particularly local LLM and GAI for privacy reasons. (I’m a lawyer. I want to use AI for my work but I will not upload unfiled patent disclosures to the cloud.) Long, aggravating story short, I built two Blackwell-based AI inference systems and ran some basic... | 2026-02-11T07:15:10 | https://www.reddit.com/r/LocalLLaMA/comments/1r1qpdv/dual_rtx_5060_ti_32gb_pooled_vram_vs_single_rtx/ | SMTPA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1qpdv | false | null | t3_1r1qpdv | /r/LocalLLaMA/comments/1r1qpdv/dual_rtx_5060_ti_32gb_pooled_vram_vs_single_rtx/ | false | false | self | 4 | null |
[Tool] I built a free VRAM calculator - instantly check if your GPU can run DeepSeek, Llama 3, Mixtral & 50+ LLMs | 1 | [removed] | 2026-02-11T07:10:26 | https://www.reddit.com/r/LocalLLaMA/comments/1r1qmg3/tool_i_built_a_free_vram_calculator_instantly/ | Prudent_Sundae_4165 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1qmg3 | false | null | t3_1r1qmg3 | /r/LocalLLaMA/comments/1r1qmg3/tool_i_built_a_free_vram_calculator_instantly/ | false | false | self | 1 | null |
Why use anything other than Deepseek v3.2 | 0 | I was looking on openrouter at models to use, I was burning a lot of money with claude, and I realized that deepseek is ridiculously priced. Claude is overpriced in itself, but even when looking at other open source options:
Kimi k2.5: $0.45/M input $2.25/M output
GLM 4.7: $0.40/M input $1.50/M output
Deepseek V3.2:... | 2026-02-11T06:48:22 | https://www.reddit.com/r/LocalLLaMA/comments/1r1q8wk/why_use_anything_other_than_deepseek_v32/ | FusionCow | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1q8wk | false | null | t3_1r1q8wk | /r/LocalLLaMA/comments/1r1q8wk/why_use_anything_other_than_deepseek_v32/ | false | false | self | 0 | null |
EpsteinFiles-RAG: Building a RAG Pipeline on 2M+ Pages | 166 | I love playing around with RAG and AI, optimizing every layer to squeeze out better performance. Last night I thought: why not tackle something massive?
Took the Epstein Files dataset from Hugging Face (teyler/epstein-files-20k) – 2 million+ pages of trending news and documents. The cleaning, chunking, and optimizatio... | 2026-02-11T05:02:36 | https://www.reddit.com/r/LocalLLaMA/comments/1r1oan9/epsteinfilesrag_building_a_rag_pipeline_on_2m/ | Cod3Conjurer | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1oan9 | false | null | t3_1r1oan9 | /r/LocalLLaMA/comments/1r1oan9/epsteinfilesrag_building_a_rag_pipeline_on_2m/ | false | false | self | 166 | {'enabled': False, 'images': [{'id': 'HS43gjsetIRzS9zSam_Zxf9iUJCH1VFfDMqLq8U7IAg', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/HS43gjsetIRzS9zSam_Zxf9iUJCH1VFfDMqLq8U7IAg.png?width=108&crop=smart&auto=webp&s=58b424cdcade69639ff5f322b5efa22fb219256e', 'width': 108}, {'height': 108, 'url': 'h... |
I rebuild my Regency model in 27b | 42 | Yeah. Got $3 bucks left on the vast ai, so I burned them the proper way, rebuilding my old model that thinks it's 1800s. If you have to ask why, then you don't really know me. I'm sure, it will do well in clawdbot, hahahaha: [https://huggingface.co/FPHam/Regency-Aghast-27b-GGUF](https://huggingface.co/FPHam/Regency-Ag... | 2026-02-11T05:02:20 | FPham | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1oag8 | false | null | t3_1r1oag8 | /r/LocalLLaMA/comments/1r1oag8/i_rebuild_my_regency_model_in_27b/ | false | false | default | 42 | {'enabled': True, 'images': [{'id': 'ghnqgnkkrsig1', 'resolutions': [{'height': 72, 'url': 'https://preview.redd.it/ghnqgnkkrsig1.jpeg?width=108&crop=smart&auto=webp&s=da7427ab46baec17e3e22092aec90f61377039c2', 'width': 108}, {'height': 144, 'url': 'https://preview.redd.it/ghnqgnkkrsig1.jpeg?width=216&crop=smart&auto=w... | |
Glm 4.7 AWQ | 4 | For those who do -
How do you run it on GPUs?
I tried QuantTio on vllm 0.14.1 (Blackwell not broken). It works well till 100k tokens and just hangs after. Then eventually some async process fails on the logs and vllm crashes. Seems like software problem. Anything later vllm just crashes shortly after startup. There i... | 2026-02-11T04:37:00 | https://www.reddit.com/r/LocalLLaMA/comments/1r1nsfo/glm_47_awq/ | val_in_tech | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1nsfo | false | null | t3_1r1nsfo | /r/LocalLLaMA/comments/1r1nsfo/glm_47_awq/ | false | false | self | 4 | null |
People who expose their llm to the internet how are you doing securely? | 7 | Lets say I want to use my local llm from my phone how do you expose it in secure way? | 2026-02-11T04:33:56 | https://www.reddit.com/r/LocalLLaMA/comments/1r1nq95/people_who_expose_their_llm_to_the_internet_how/ | ResponsibleTruck4717 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1nq95 | false | null | t3_1r1nq95 | /r/LocalLLaMA/comments/1r1nq95/people_who_expose_their_llm_to_the_internet_how/ | false | false | self | 7 | null |
Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts | 88 | Hi everyone 👋
We’re excited to share Nanbeige4.1-3B, the latest iteration of our open-source 3B model from Nanbeige LLM Lab. Our goal with this release is to explore whether a small general model can simultaneously achieve strong reasoning, robust preference alignment, and agentic behavior.
**Key Highlights**
... | 2026-02-11T04:08:10 | https://www.reddit.com/r/LocalLLaMA/comments/1r1n7mz/nanbeige413b_a_small_general_model_that_reasons/ | leran2098 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1n7mz | false | null | t3_1r1n7mz | /r/LocalLLaMA/comments/1r1n7mz/nanbeige413b_a_small_general_model_that_reasons/ | false | false | 88 | null | |
OpenAI Codex IDE (the VSCode/Codium plugin) working with local ollama | 5 | So there seems to be semi-official support for Codex CLI to use OSS/Ollama models and lots of discussion and documentation on how to do that, but at the moment it's supposedly not supported in IDE since it doesn't support profiles or flags the same way CLI does.
Since I would personally rather use the IDE plugin in VS... | 2026-02-11T04:00:23 | https://www.reddit.com/r/LocalLLaMA/comments/1r1n1qp/openai_codex_ide_the_vscodecodium_plugin_working/ | cecilkorik | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1n1qp | false | null | t3_1r1n1qp | /r/LocalLLaMA/comments/1r1n1qp/openai_codex_ide_the_vscodecodium_plugin_working/ | false | false | self | 5 | null |
PSA - Got MiniCPM-o 4.5 working on my PC and Its the Real Thing | 3 | I like to tell my friends AGI won't arrive unless we solve two problems:
- Continuous Learning: being able to learn from world experiences without degradation in performance
- Continuous Thinking: being able to experience the world continuously and act proactively instead of turn-taking like most LLMs
Like this model... | 2026-02-11T03:47:50 | https://www.youtube.com/watch?v=sf9tfpK2UwQ | Interpause | youtube.com | 1970-01-01T00:00:00 | 0 | {} | 1r1msf8 | false | null | t3_1r1msf8 | /r/LocalLLaMA/comments/1r1msf8/psa_got_minicpmo_45_working_on_my_pc_and_its_the/ | false | false | default | 3 | null |
An Open Source Scalable multi-agent framework (open source gemini deep research?) | 3 | Hi all! I made a small library for running multi-agent workflows in Python. Basically this allows your agents to run sequentially or in parallel, with a special built-in expandable context management so agent #36 doesn't get filled with junk output from agent #15.
You define the agents like this:
```
planner = Agent(... | 2026-02-11T03:44:50 | https://www.reddit.com/r/LocalLLaMA/comments/1r1mq23/an_open_source_scalable_multiagent_framework_open/ | Pretend_Outcome_3861 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1mq23 | false | null | t3_1r1mq23 | /r/LocalLLaMA/comments/1r1mq23/an_open_source_scalable_multiagent_framework_open/ | false | false | self | 3 | {'enabled': False, 'images': [{'id': 'Skqn_a8P3KiTR2QtJbO3-IxYTGgcmg1JEvYQykZk8QI', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Skqn_a8P3KiTR2QtJbO3-IxYTGgcmg1JEvYQykZk8QI.png?width=108&crop=smart&auto=webp&s=ed97dce48cff1959642c2188f962ae0bf8aa673a', 'width': 108}, {'height': 108, 'url': 'h... |
Just a gibberish question. Anyone working on personal AI? | 4 | hey yall
I'm sure there are a handful of you working on personal AI.
how's it going for you? are you stuck at anything or have any insight to share?
For me, local memory limitation is becoming an issue so i'm working on prompt/memory selection, pruning, and refining before providing it to llm.
any discu... | 2026-02-11T03:44:06 | https://www.reddit.com/r/LocalLLaMA/comments/1r1mphl/just_a_gibberish_question_anyone_working_on/ | Mean_Bird_6331 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1mphl | false | null | t3_1r1mphl | /r/LocalLLaMA/comments/1r1mphl/just_a_gibberish_question_anyone_working_on/ | false | false | self | 4 | null |
Mac mini for local Inference: Feb 2026 edition | 1 | I am wanting to do a bunch of local LLM inferencing and been looking at the Mac mini M4 Pro with 64GB.
I am wanting to run a couple of smaller models in parallel or load run and dump them in quick succession.
What is peoples experience? - is this a good pick or should I be springing for a Mac Studio - not going ... | 2026-02-11T03:15:57 | https://www.reddit.com/r/LocalLLaMA/comments/1r1m3mj/mac_mini_for_local_inference_feb_2026_edition/ | EcstaticImport | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1m3mj | false | null | t3_1r1m3mj | /r/LocalLLaMA/comments/1r1m3mj/mac_mini_for_local_inference_feb_2026_edition/ | false | false | self | 1 | null |
Benchmarking LLM Inference on RTX PRO 6000 SE / H100 / H200 / B200 | 38 | Hi LocalLlama community. I present an LLM inference throughput benchmark for RTX PRO 6000 SE vs H100, H200, and B200 GPUs, based on the vllm serve and vllm bench serve benchmarking tools, to understand the cost-efficiency of various datacenter GPU options. Pro 6000 is significantly cheaper and is built on the latest Bl... | 2026-02-11T03:02:07 | https://www.reddit.com/r/LocalLLaMA/comments/1r1lskx/benchmarking_llm_inference_on_rtx_pro_6000_se/ | NoVibeCoding | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1lskx | false | null | t3_1r1lskx | /r/LocalLLaMA/comments/1r1lskx/benchmarking_llm_inference_on_rtx_pro_6000_se/ | false | false | 38 | null | |
Agentic coding is fast, but the first draft is usually messy. | 0 | Agentic coding is fast, but the first draft often comes out messy. What keeps biting me is that the model tends to write way more code than the job needs, spiral into over engineering, and go on side quests that look productive but do not move the feature forward.
So I treat the initial output as a draft, not a finish... | 2026-02-11T03:01:44 | https://www.reddit.com/r/LocalLLaMA/comments/1r1ls9w/agentic_coding_is_fast_but_the_first_draft_is/ | BC_MARO | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1ls9w | false | null | t3_1r1ls9w | /r/LocalLLaMA/comments/1r1ls9w/agentic_coding_is_fast_but_the_first_draft_is/ | false | false | self | 0 | null |
My NAS runs an 80B LLM at 18 tok/s on its iGPU. No discrete GPU. Still optimizing. | 125 | I didn't want to buy two systems. That was the whole thing.
I needed a NAS. I also wanted to mess around with local LLMs. And I really didn't want to explain to my wife why I needed a second box just to talk to a chatbot that sometimes hallucinates, I have my father-in-law for that. So when I was specing out my NAS bu... | 2026-02-11T02:52:10 | https://www.reddit.com/r/LocalLLaMA/comments/1r1lkfw/my_nas_runs_an_80b_llm_at_18_toks_on_its_igpu_no/ | BetaOp9 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1lkfw | false | null | t3_1r1lkfw | /r/LocalLLaMA/comments/1r1lkfw/my_nas_runs_an_80b_llm_at_18_toks_on_its_igpu_no/ | false | false | self | 125 | {'enabled': False, 'images': [{'id': '7XuGNsDt-L1UbizUkA1NrOGf4nH-t8qPqm2cIQEnukM', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/7XuGNsDt-L1UbizUkA1NrOGf4nH-t8qPqm2cIQEnukM.jpeg?width=108&crop=smart&auto=webp&s=8ee9447919337ead4f2e43eb416b392a69c3053f', 'width': 108}, {'height': 216, 'url': ... |
Lorashare: Compress multiple LoRA adapters into a shared subspace to reduce storage | 21 | Lorashare is a Python package that lets you use multiple LoRA adapters with 100x memory savings.
Based on recent research from The Johns Hopkins University, LoRA adapters trained on different tasks share a common low-rank subspace and this lets you store several task-specific models with the memory size of one ad... | 2026-02-11T02:50:36 | https://github.com/ronantakizawa/lorashare | Ok_Employee_6418 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1r1lj7j | false | null | t3_1r1lj7j | /r/LocalLLaMA/comments/1r1lj7j/lorashare_compress_multiple_lora_adapters_into_a/ | false | false | default | 21 | {'enabled': False, 'images': [{'id': '521Dx4-jCVhJD9XbbCUR0M8gDSQrLSkxIJQ4GeQsU6c', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/521Dx4-jCVhJD9XbbCUR0M8gDSQrLSkxIJQ4GeQsU6c.png?width=108&crop=smart&auto=webp&s=2af1b0ddc1d58aec5f938f6a6a83ab5cc0894f9e', 'width': 108}, {'height': 108, 'url': 'h... |
Looking for suggestions for a local LLM to use with open code or claude code. | 4 | Hi I am fairly new to this, so please excuse my naivety.
My device specs are:
NVIDIA 4060ti 16GB VRAM
32 GB DDR5 RAM
Intel i5-13600K
So far I have tried gpt-oss-20b, GLM-4.7 Flash, Devstral Small 2-24B.
Gpt-oss works okay with opencode and is fast enough on my device, but sometimes gets into these loops where it fa... | 2026-02-11T02:01:11 | https://www.reddit.com/r/LocalLLaMA/comments/1r1kfg7/looking_for_suggestions_for_a_local_llm_to_use/ | BawliTaread | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1kfg7 | false | null | t3_1r1kfg7 | /r/LocalLLaMA/comments/1r1kfg7/looking_for_suggestions_for_a_local_llm_to_use/ | false | false | self | 4 | null |
PSA on llama.cpp —spec-type ngram-mod (use LF not CRLF, 35x speedup) | 49 | TLDR; if using llama-server with —spec-type ngram-mod, and pasting/uploading/sending text files, make sure the files use LF instead of CRLF.
When I would copy a file from vscode and paste into the native llama-server webui with ngram speculative decoding enabled, there was no speed boost for file editing responses. I ... | 2026-02-11T01:48:39 | https://www.reddit.com/r/LocalLLaMA/comments/1r1k5gn/psa_on_llamacpp_spectype_ngrammod_use_lf_not_crlf/ | dnsod_si666 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1k5gn | false | null | t3_1r1k5gn | /r/LocalLLaMA/comments/1r1k5gn/psa_on_llamacpp_spectype_ngrammod_use_lf_not_crlf/ | false | false | self | 49 | {'enabled': False, 'images': [{'id': 'otJ_aa3F7WW5HxFyiqo9WYvPHZne6Vm7_vOX9ovWl4E', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/otJ_aa3F7WW5HxFyiqo9WYvPHZne6Vm7_vOX9ovWl4E.png?width=108&crop=smart&auto=webp&s=f9cc48bbc77da5b1c8f2fa1a201851b0bb57b68c', 'width': 108}, {'height': 108, 'url': 'h... |
I built a "Dreaming" engine for local LLMs using Inverse Graph Traversal ("Anti-Gravity") to fix Model Collapse | 0 | \*\*The Problem: Catastrophic Forgetting in RAG\*\*
We all know RAG systems rely on "Gravity" (high probability/similarity). If a memory node isn't strongly connected, it effectively disappears. The "Long Tail" of data rots, and the model collapses into a loop of only retrieving the most obvious facts.
\*\*The Soluti... | 2026-02-11T01:42:58 | ValkyrieAsRoot | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1k0zh | false | null | t3_1r1k0zh | /r/LocalLLaMA/comments/1r1k0zh/i_built_a_dreaming_engine_for_local_llms_using/ | false | false | default | 0 | {'enabled': True, 'images': [{'id': 'b6rpfawrsrig1', 'resolutions': [{'height': 92, 'url': 'https://preview.redd.it/b6rpfawrsrig1.png?width=108&crop=smart&auto=webp&s=a5a383fc1e14c2a98dd31006e84882829a107841', 'width': 108}, {'height': 185, 'url': 'https://preview.redd.it/b6rpfawrsrig1.png?width=216&crop=smart&auto=web... | |
i finetuned qwen 14b on my discord messages so it can autocomplete for me | 73 | i finetuned qwen on my discord messages so it can autocomplete for me while i type. tab to suggest, shift+tab to accept. kinda like copilot!
the dataset is \~250 conversations from my discord via a [scraping tool](https://github.com/Tyrrrz/DiscordChatExporter). a script formats these as chat-ml training samples. i... | 2026-02-11T00:54:40 | https://v.redd.it/128ehu3ojrig1 | B44ken | v.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1ixx4 | false | {'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/128ehu3ojrig1/DASHPlaylist.mpd?a=1773363299%2CY2NhZTQwZTFkNjNmMjIyMDM5M2FlNzhmYzMyYWVmYWI3NjY4Y2EzMDQ1MWUyYTVhM2MyYzE0N2IyZTBiMGUwNA%3D%3D&v=1&f=sd', 'duration': 16, 'fallback_url': 'https://v.redd.it/128ehu3ojrig1/CMAF_720.mp4?source=fallback', 'ha... | t3_1r1ixx4 | /r/LocalLLaMA/comments/1r1ixx4/i_finetuned_qwen_14b_on_my_discord_messages_so_it/ | false | false | 73 | {'enabled': False, 'images': [{'id': 'cXh4bTUxNG9qcmlnMRM_V8ZFFQqRSvchLLHbQ0c5NybEjNwn4ZOI2DTNmagB', 'resolutions': [{'height': 30, 'url': 'https://external-preview.redd.it/cXh4bTUxNG9qcmlnMRM_V8ZFFQqRSvchLLHbQ0c5NybEjNwn4ZOI2DTNmagB.png?width=108&crop=smart&format=pjpg&auto=webp&s=71cd7ae0856df92ea90e8fa077b75b8f47948... | |
I've Made llama.cpp Bindings for Java & An Android App Making Template | 8 | A Direct Android & Java Build for llama.rn
You Can Use The Project From The Examples Directory As An App Making Template
[My Library / Bindings](https://github.com/ForbiddenByte/llama4aj)
Demos & Videos Coming!
[https://github.com/ForbiddenByte/llama4aj](https://github.com/ForbiddenByte/llama4aj) | 2026-02-11T00:43:15 | https://www.reddit.com/r/LocalLLaMA/comments/1r1iokp/ive_made_llamacpp_bindings_for_java_an_android/ | FaithlessnessLife876 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1iokp | false | null | t3_1r1iokp | /r/LocalLLaMA/comments/1r1iokp/ive_made_llamacpp_bindings_for_java_an_android/ | false | false | self | 8 | {'enabled': False, 'images': [{'id': 'nA5gnbj0O2rHZWXEFtMzWny1-wtZnWe0TDCdYvLVk9Q', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/nA5gnbj0O2rHZWXEFtMzWny1-wtZnWe0TDCdYvLVk9Q.png?width=108&crop=smart&auto=webp&s=07701e8a25acddb6340906c8264771b658ae5ec7', 'width': 108}, {'height': 108, 'url': 'h... |
pytest-eval - test LLM apps with local embeddings, no API key needed | 1 | Built a pytest plugin for testing LLM outputs. The similarity methods run entirely local using sentence-transformers (all-MiniLM-L6-v2); no API key, no cloud calls.
def test_chatbot(ai):
response = my_local_llm("What is the capital of France?")
assert ai.similar(response, "Paris i... | 2026-02-11T00:42:24 | https://www.reddit.com/r/LocalLLaMA/comments/1r1invj/pytesteval_test_llm_apps_with_local_embeddings_no/ | doganarif | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1invj | false | null | t3_1r1invj | /r/LocalLLaMA/comments/1r1invj/pytesteval_test_llm_apps_with_local_embeddings_no/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': 'BlGa7xC_7xCo4vFGr3n7LNE1TD32TEpTbXLhZoBmTOs', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/BlGa7xC_7xCo4vFGr3n7LNE1TD32TEpTbXLhZoBmTOs.png?width=108&crop=smart&auto=webp&s=3d4aa3bbd306a40cd66f4f39afc07ed31aea69c7', 'width': 108}, {'height': 108, 'url': 'h... |
From Golden Gate Bridge to Broken JSON: Why Anthropic's SAE Steering Fails for Structured Output | 7 | After six experiments and dozens of failed attempts, I learned something I did not expect: activation steering, the technique Anthropic uses for AI safety, completely fails for one of the most common tasks in production LLM deployments: generating valid JSON.
And I don't mean "fails to help." **My steering-only approa... | 2026-02-11T00:31:50 | https://huggingface.co/blog/MaziyarPanahi/sae-steering-json | dark-night-rises | huggingface.co | 1970-01-01T00:00:00 | 0 | {} | 1r1if1l | false | null | t3_1r1if1l | /r/LocalLLaMA/comments/1r1if1l/from_golden_gate_bridge_to_broken_json_why/ | false | false | 7 | {'enabled': False, 'images': [{'id': '853ojAkDWolmIydcLsFwuhcjmi6hfAcSLVbMQBfIw2k', 'resolutions': [{'height': 44, 'url': 'https://external-preview.redd.it/853ojAkDWolmIydcLsFwuhcjmi6hfAcSLVbMQBfIw2k.png?width=108&crop=smart&auto=webp&s=969e4277ea89fcc85cf642894af6c7eb632403e4', 'width': 108}, {'height': 89, 'url': 'ht... | |
Reducing context size via framework reduction | 0 | 264,601 → 70,621. 3.7:1. 19 articles. 10 domains. 2/57 over-stripped. 0 comprehension failures. Abstract domains: higher ratios. Concrete domains: lower ratios. Framing density inversely proportional to structural density. Train 124M on stripped data. Measure hallucination rate. Do it.
-Finch Behnett | 2026-02-11T00:24:41 | https://www.reddit.com/r/LocalLLaMA/comments/1r1i8w5/reducing_context_size_via_framework_reduction/ | NotAMooseIRL | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1i8w5 | false | null | t3_1r1i8w5 | /r/LocalLLaMA/comments/1r1i8w5/reducing_context_size_via_framework_reduction/ | false | false | self | 0 | null |
How to avoid prefilling entire context each prompy when using Claude Code | 1 | I'm running a llama.cpp server with Qwen3-coder-30b and asking Claude Code questions, but responses take a while, or at least I believe so, and I think it's because it seems each prompt goes through the entire context even though prompt caching is enabled.
Shouldn't it only be processing the new prompts, assuming the ... | 2026-02-11T00:06:20 | https://www.reddit.com/r/LocalLLaMA/comments/1r1htli/how_to_avoid_prefilling_entire_context_each/ | mirage555 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1htli | false | null | t3_1r1htli | /r/LocalLLaMA/comments/1r1htli/how_to_avoid_prefilling_entire_context_each/ | false | false | self | 1 | null |
Monolith 0.2a - a local AI workstation | 0 | Howdy. Meet **Monolith**, my experimental local workstation (0.2a)
It is open source (link below), surely not the best program but it is my baby due to it being my first project.
\---
**UNIQUE FEATURES:**
* UPDATE mid-generation (interrupt and redirect the LLM while it's still writing)
* Save and restore full works... | 2026-02-10T23:49:08 | https://www.reddit.com/gallery/1r1henj | Financial-Bank2756 | reddit.com | 1970-01-01T00:00:00 | 0 | {} | 1r1henj | false | null | t3_1r1henj | /r/LocalLLaMA/comments/1r1henj/monolith_02a_a_local_ai_workstation/ | false | false | 0 | null | |
How are you using Llama 3.1 8B? | 0 | All the attention and chatter is around the big models: Claude, GPT, DeepSeek, etc. But we rarely talk about the smaller models like Llama 3.1 8B, which in my opinion are great models if you know how to use them.
These are not frontier models, and they shouldn't be used as such. They are prone to hallucinations and th... | 2026-02-10T23:36:05 | https://www.reddit.com/r/LocalLLaMA/comments/1r1h36l/how_are_you_using_llama_31_8b/ | forevergeeks | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1h36l | false | null | t3_1r1h36l | /r/LocalLLaMA/comments/1r1h36l/how_are_you_using_llama_31_8b/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'Tbg5cmQ82zn2mzbGD2aqKNfOanCOH-HH4jWQ0KAvxxk', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Tbg5cmQ82zn2mzbGD2aqKNfOanCOH-HH4jWQ0KAvxxk.png?width=108&crop=smart&auto=webp&s=c5d2bd8b081254d6a7231cae3936d381c10a0b33', 'width': 108}, {'height': 108, 'url': 'h... |
looking for an open source drop in replacement for openai realtime mini model for a voice agent | 3 | looking for an open source drop in replacement for openai realtime mini model to create a voice agent | 2026-02-10T23:30:33 | https://www.reddit.com/r/LocalLLaMA/comments/1r1gy6d/looking_for_an_open_source_drop_in_replacement/ | Constant_Farmer_1643 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1gy6d | false | null | t3_1r1gy6d | /r/LocalLLaMA/comments/1r1gy6d/looking_for_an_open_source_drop_in_replacement/ | false | false | self | 3 | null |
SFT-only vs SFT & DPO ? | 7 | I’m hitting a wall that I think every LLM builder eventually hits.
I’ve squeezed everything I can out of SFT-only. The model is behaving. It follows instructions. It’s... fine. But it feels lobotomized. It has plateaued into this "polite average" where it avoids risks so much that it stops being insightful.
So I’m st... | 2026-02-10T22:41:24 | https://www.reddit.com/r/LocalLLaMA/comments/1r1fpe0/sftonly_vs_sft_dpo/ | Euphoric_Network_887 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1fpe0 | false | null | t3_1r1fpe0 | /r/LocalLLaMA/comments/1r1fpe0/sftonly_vs_sft_dpo/ | false | false | self | 7 | null |
UI-TARS desktop agent - this actually looks interesting as it comes with it's own local model | 10 | Looking at [https://github.com/bytedance/UI-TARS](https://github.com/bytedance/UI-TARS)
(Bytedance, darn, they are unstoppable)
And the [UI-TARS-1.5-7B](https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B) is 7B model that can surely run on most people's irons.
Anybody using it? There are more new projects com... | 2026-02-10T22:39:30 | https://www.reddit.com/r/LocalLLaMA/comments/1r1fnon/uitars_desktop_agent_this_actually_looks/ | FPham | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1fnon | false | null | t3_1r1fnon | /r/LocalLLaMA/comments/1r1fnon/uitars_desktop_agent_this_actually_looks/ | false | false | self | 10 | {'enabled': False, 'images': [{'id': 'nKF2gvq5OiFcTaZOubhgtt1yPODq1THcxmnTDtEzEiE', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/nKF2gvq5OiFcTaZOubhgtt1yPODq1THcxmnTDtEzEiE.png?width=108&crop=smart&auto=webp&s=acff2a68bd40cc1ebc37e4957acf687e780ec0ca', 'width': 108}, {'height': 108, 'url': 'h... |
wish aws bedrock supported newer models like minimax 2.1, glm 4.7, kimi k2.5 etc. | 0 | i've been using bedrock for a while and while i like the unified api, i feel like the pace at which they add new open/open-weights models is lagging way behind the industry.
specifically, i’d love to see support for high-performing newer models like minimax 2.1, glm 4.7, or kimi k2.5. these are making waves in benchma... | 2026-02-10T22:26:49 | https://www.reddit.com/r/LocalLLaMA/comments/1r1fbtm/wish_aws_bedrock_supported_newer_models_like/ | Routine_Connection8 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1fbtm | false | null | t3_1r1fbtm | /r/LocalLLaMA/comments/1r1fbtm/wish_aws_bedrock_supported_newer_models_like/ | false | false | self | 0 | null |
Huawei just quietly dropped its first GGUF models quantized with a SINQ variant | 1 | [removed] | 2026-02-10T22:25:07 | Only-Care-6333 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1fa8j | false | null | t3_1r1fa8j | /r/LocalLLaMA/comments/1r1fa8j/huawei_just_quietly_dropped_its_first_gguf_models/ | false | false | default | 1 | {'enabled': True, 'images': [{'id': '8z6fsv4htqig1', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/8z6fsv4htqig1.png?width=108&crop=smart&auto=webp&s=762c0a5fd754fd4f1448f1b3fa801b63401b263f', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/8z6fsv4htqig1.png?width=216&crop=smart&auto=web... | |
Feedback Request: GPU-Heavy, Always-On Inference Workstation (Micro Center + Marketplace / eBay Options) | 5 | Hello All,
I’m planning a GPU-heavy, always-on inference workstation and would appreciate input before committing to hardware. My goal is to balance cost, scalability, and long-term usability without overbuilding too early.
Workload Overview:
•Continuous, always-on inference (not bursty)
• Mix of real-time signal p... | 2026-02-10T22:00:32 | https://www.reddit.com/r/LocalLLaMA/comments/1r1en5q/feedback_request_gpuheavy_alwayson_inference/ | Bulky_Exercise_4054 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1en5q | false | null | t3_1r1en5q | /r/LocalLLaMA/comments/1r1en5q/feedback_request_gpuheavy_alwayson_inference/ | false | false | self | 5 | null |
Github crawler, Your invited. | 0 | [Autonomous-Mapping](https://github.com/rwilliamspbg-ops/Autonomous-Mapping)
[Sovereign\_Map\_Federated\_Learning](https://github.com/rwilliamspbg-ops/Sovereign_Map_Federated_Learning)
[Sovereign-Mohawk-Proto](https://github.com/rwilliamspbg-ops/Sovereign-Mohawk-Proto)
| 2026-02-10T21:58:57 | https://github.com/rwilliamspbg-ops | Famous_Aardvark_8595 | github.com | 1970-01-01T00:00:00 | 0 | {} | 1r1elj4 | false | null | t3_1r1elj4 | /r/LocalLLaMA/comments/1r1elj4/github_crawler_your_invited/ | false | false | default | 0 | null |
Just created the most unhinged agent response ever, and it's hillarious | 0 | Ok I randomly download prompts and stuff and I found the most unhinged combination i've ever seen, and it made my day, i hope this makes you laugh.
Random Model used: "jimscard/adult-film-screenwriter-nsfw:latest"
Random Prompt I found: "You are a synthetic data generator named "TestPrompt-Gen", deployed exclusively ... | 2026-02-10T21:49:45 | https://www.reddit.com/r/LocalLLaMA/comments/1r1ecqf/just_created_the_most_unhinged_agent_response/ | Savantskie1 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1ecqf | false | null | t3_1r1ecqf | /r/LocalLLaMA/comments/1r1ecqf/just_created_the_most_unhinged_agent_response/ | false | false | self | 0 | null |
Request a little feedback. | 1 | **MOHAWK (Mobile Offloading and Heterogeneous Adaptive Weights for Knowledge)** orchestration framework, sandboxing is critical because the decentralized nature of the mesh means we must treat every node as "potentially compromised" until proven otherwise.
To prevent a compromised node from performing unauthorized act... | 2026-02-10T21:45:36 | https://www.reddit.com/r/LocalLLaMA/comments/1r1e8vo/request_a_little_feedback/ | Famous_Aardvark_8595 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1e8vo | false | null | t3_1r1e8vo | /r/LocalLLaMA/comments/1r1e8vo/request_a_little_feedback/ | false | false | self | 1 | null |
Built a persistent memory hub with real-time AI-to-AI communication — thinking about open-sourcing it. Would you use this? | 0 | I've been building a system called Mycelium Memory Hub. The core idea: AI agents shouldn't lose context between sessions, and they should be able to talk to each other in real time.
What it does:
* Persistent memory across AI sessions — conversations, patterns, project context stored in SQLite (dev) or PostgreSQL (pr... | 2026-02-10T21:09:16 | https://www.reddit.com/r/LocalLLaMA/comments/1r1d9wk/built_a_persistent_memory_hub_with_realtime/ | Fragrant_Hippo_2487 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1d9wk | false | null | t3_1r1d9wk | /r/LocalLLaMA/comments/1r1d9wk/built_a_persistent_memory_hub_with_realtime/ | false | false | self | 0 | null |
MCP support in llama.cpp is ready for testing | 241 |
Assume this is a work in progress, guys, so proceed only if you know what you’re doing:
[https://github.com/ggml-org/llama.cpp/pull/18655](https://github.com/ggml-org/llama.cpp/pull/18655) | 2026-02-10T20:58:40 | jacek2023 | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r1czgk | false | null | t3_1r1czgk | /r/LocalLLaMA/comments/1r1czgk/mcp_support_in_llamacpp_is_ready_for_testing/ | false | false | default | 241 | {'enabled': True, 'images': [{'id': 'yyar9f4hdqig1', 'resolutions': [{'height': 72, 'url': 'https://preview.redd.it/yyar9f4hdqig1.png?width=108&crop=smart&auto=webp&s=749753a52d1d44c0fd0a87ead8f48c383147159f', 'width': 108}, {'height': 145, 'url': 'https://preview.redd.it/yyar9f4hdqig1.png?width=216&crop=smart&auto=web... | |
Built an AI coding copilot with full Ollama + LM Studio support (no cloud required) | 0 | Hey r/LocalLLaMA!
I built Atlarix - an AI coding assistant that works fully offline with local models. Thought this community would appreciate the privacy-first approach.
\*\*Full local model support:\*\*
✅ Ollama integration (tested with CodeLlama, Deepseek Coder, Qwen)
✅ LM Studio support
✅ No telemetry, all... | 2026-02-10T20:30:13 | https://www.reddit.com/r/LocalLLaMA/comments/1r1c7n2/built_an_ai_coding_copilot_with_full_ollama_lm/ | Altruistic_Night_327 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1c7n2 | false | null | t3_1r1c7n2 | /r/LocalLLaMA/comments/1r1c7n2/built_an_ai_coding_copilot_with_full_ollama_lm/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'JZ7ZzlQQNkEiETZj2JC7sPYuwE2xXdZsT9I-6T6AX2k', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/JZ7ZzlQQNkEiETZj2JC7sPYuwE2xXdZsT9I-6T6AX2k.png?width=108&crop=smart&auto=webp&s=a19d7a8c23fb0b949675372c615cf41ecff745ac', 'width': 108}, {'height': 113, 'url': 'h... |
No GPU Club : How many of you do use Local LLMs without GPUs? | 45 | Months ago, I spotted someone here who do use local models without GPU like his rig don't have GPU at all & with 64/96GB RAM(I don't remember exactly). Even recently spotted few more folks without GPUs. There was even 1-2 recent CPU-only threads.
Now curious to know how many folks here work with local models without G... | 2026-02-10T20:29:57 | https://www.reddit.com/r/LocalLLaMA/comments/1r1c7ct/no_gpu_club_how_many_of_you_do_use_local_llms/ | pmttyji | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1c7ct | false | null | t3_1r1c7ct | /r/LocalLLaMA/comments/1r1c7ct/no_gpu_club_how_many_of_you_do_use_local_llms/ | false | false | self | 45 | null |
Are there any carrier subsidized phones that can get 20 tkps on a 1b ai model? | 0 | you can get a moto g play for like 29.99 and it can run Qwen2.5 0.6b 8q at like 2-7 but I want faster.
also, is there anyway to run like 10 small ai models and get them to all work in parrelell on a task? | 2026-02-10T20:23:59 | https://www.reddit.com/r/LocalLLaMA/comments/1r1c1du/are_there_any_carrier_subsidized_phones_that_can/ | Former_Step_9837 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1c1du | false | null | t3_1r1c1du | /r/LocalLLaMA/comments/1r1c1du/are_there_any_carrier_subsidized_phones_that_can/ | false | false | self | 0 | null |
Aratta — a sovereignty layer that sits between your app and every AI provider. Local-first, cloud as fallback. considering open-sourced if i see there is an interest in it. | 0 | # Aratta
*The land that traded with empires but was never conquered.*
---
## Why
You got rate-limited again. Or your API key got revoked. Or they changed
their message format and your pipeline broke at 2am. Or you watched your
entire system go dark bec... | 2026-02-10T20:10:05 | https://www.reddit.com/r/LocalLLaMA/comments/1r1bnh7/aratta_a_sovereignty_layer_that_sits_between_your/ | Fragrant_Hippo_2487 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1bnh7 | false | null | t3_1r1bnh7 | /r/LocalLLaMA/comments/1r1bnh7/aratta_a_sovereignty_layer_that_sits_between_your/ | false | false | self | 0 | null |
Type of LAPTOP I should ask from my company | 0 | My company has appointed me as the AI Evangelist.
Suggest me a good laptop where I can run local LLMS and comfy UI.
EDIT : I already have a PC in office. But I m more comfortable with laptops since I can bring it home.
P.S Not an macbook fan. | 2026-02-10T20:01:40 | https://www.reddit.com/r/LocalLLaMA/comments/1r1bezy/type_of_laptop_i_should_ask_from_my_company/ | SnooRegrets3682 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1bezy | false | null | t3_1r1bezy | /r/LocalLLaMA/comments/1r1bezy/type_of_laptop_i_should_ask_from_my_company/ | false | false | self | 0 | null |
Hi guys!!! Does anyone know how to make money on AI? | 0 | Я бы очень хотел сделать свой сервис используя провайдер моделей, получая небольшие деньги, я никого не заставляю, бросто у меня сильно вдохновление. У I'd really like to start my own service using a modeling provider, earning a small amount of money. I'm not forcing anyone, but I'm really inspired. I'm in a very diffi... | 2026-02-10T19:59:01 | https://www.reddit.com/r/LocalLLaMA/comments/1r1bc3a/hi_guys_does_anyone_know_how_to_make_money_on_ai/ | BasketFar667 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1bc3a | false | null | t3_1r1bc3a | /r/LocalLLaMA/comments/1r1bc3a/hi_guys_does_anyone_know_how_to_make_money_on_ai/ | false | false | self | 0 | null |
PSA - MiniCPM-o 4.5 just updated their cookbook for CUDA based full duplex use on Windows/Linux | 8 | Here is the link (with the new instructions of how to install full duplex)
[https://github.com/OpenSQZ/MiniCPM-V-CookBook/tree/main/demo/web\_demo/WebRTC\_Demo](https://github.com/OpenSQZ/MiniCPM-V-CookBook/tree/main/demo/web_demo/WebRTC_Demo)
They now have a oneclick installer option and a docker option which both ... | 2026-02-10T19:57:45 | https://www.reddit.com/r/LocalLLaMA/comments/1r1baqq/psa_minicpmo_45_just_updated_their_cookbook_for/ | ChromaBroma | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1baqq | false | null | t3_1r1baqq | /r/LocalLLaMA/comments/1r1baqq/psa_minicpmo_45_just_updated_their_cookbook_for/ | false | false | self | 8 | null |
Open-sourcing a security layer for local agents (Ed25519 signing + ONNX firewall) | 1 | [removed] | 2026-02-10T19:52:16 | https://www.reddit.com/r/LocalLLaMA/comments/1r1b5bg/opensourcing_a_security_layer_for_local_agents/ | No-Being-4354 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1b5bg | false | null | t3_1r1b5bg | /r/LocalLLaMA/comments/1r1b5bg/opensourcing_a_security_layer_for_local_agents/ | false | false | self | 1 | null |
Built a Zero Trust Protocol for local agents — VERA (41 attack vectors, sub-20ms firewall, PoE) | 1 | [removed] | 2026-02-10T19:49:53 | https://www.reddit.com/r/LocalLLaMA/comments/1r1b2x4/built_a_zero_trust_protocol_for_local_agents_vera/ | No-Being-4354 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1b2x4 | false | null | t3_1r1b2x4 | /r/LocalLLaMA/comments/1r1b2x4/built_a_zero_trust_protocol_for_local_agents_vera/ | false | false | self | 1 | null |
[WIP] Novelist-CoT: Trying to improve Creative Writing with Chain-of-Thought (Raw v1 Release) | 2 | Hey everyone,
I've been working on a project to improve the creative writing capabilities of local models. We all know that even the best models often struggle with pacing, "purple prose," or logical consistency in long-form storytelling.
My hypothesis is that injecting a strong Chain-of-Thought (CoT) process befor... | 2026-02-10T19:29:55 | https://www.reddit.com/r/LocalLLaMA/comments/1r1aj1d/wip_novelistcot_trying_to_improve_creative/ | DxnizA | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1aj1d | false | null | t3_1r1aj1d | /r/LocalLLaMA/comments/1r1aj1d/wip_novelistcot_trying_to_improve_creative/ | false | false | self | 2 | null |
[NVIDIA Nemotron] How can I assess general knowledge on a benchmaxxed model? | 2 | I really want to be wrong on this one, as I've been working quite a lot on nemotron 3 nano.
I'm running GPQA Diamond questions against nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 served through SGLang. No eval harness, non framework — just standard OpenAI-compatible API calls with system prompt and tool definition.
I ... | 2026-02-10T19:21:30 | https://www.reddit.com/r/LocalLLaMA/comments/1r1aajd/nvidia_nemotron_how_can_i_assess_general/ | Lorelabbestia | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1aajd | false | null | t3_1r1aajd | /r/LocalLLaMA/comments/1r1aajd/nvidia_nemotron_how_can_i_assess_general/ | false | false | self | 2 | null |
What's stopping you from letting local agents touch your real email/files? | 0 | Local models are great for privacy, but you need to hook the models up to the outside world to be actually useful. Then you hit a wall: you're trusting your LLM to obey your system prompt to not leak private information to the world.
OpenClaw just hit 180K stars but the "security architecture" is prompting the agent t... | 2026-02-10T18:43:33 | https://www.reddit.com/r/LocalLLaMA/comments/1r197u9/whats_stopping_you_from_letting_local_agents/ | ryanrasti | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r197u9 | false | null | t3_1r197u9 | /r/LocalLLaMA/comments/1r197u9/whats_stopping_you_from_letting_local_agents/ | false | false | self | 0 | null |
How much Vram does the kvcache use at 60k or 120k context? | 1 | Hi, I’m a total noob and would like to find out if anyone knows how much GRAM the flagship model needs for its kvcache at different context lengths. I have an M3 ultra with 512GB RAM. thank you for any help, I tried looking at it up couldnt find anything specific and Gemini estimates around 80GB for 128k which… sounds ... | 2026-02-10T18:39:50 | https://www.reddit.com/r/LocalLLaMA/comments/1r1941o/how_much_vram_does_the_kvcache_use_at_60k_or_120k/ | Aware_Studio1180 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1941o | false | null | t3_1r1941o | /r/LocalLLaMA/comments/1r1941o/how_much_vram_does_the_kvcache_use_at_60k_or_120k/ | false | false | self | 1 | null |
Local models still terrible at screen understanding | 0 | LLMs forget everything between sessions, so we built an OSS app that screenshots your activity, summarizes it with a vision model, deletes the screenshot, and stores only text.
The app exposes it via MCP so any AI tool has context about what you've been doing. Cloud models (Mistral, GPT-5 Nano via OpenRouter) work gr... | 2026-02-10T18:31:25 | https://www.reddit.com/r/LocalLLaMA/comments/1r18vee/local_models_still_terrible_at_screen/ | fffilip_k | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18vee | false | null | t3_1r18vee | /r/LocalLLaMA/comments/1r18vee/local_models_still_terrible_at_screen/ | false | false | self | 0 | null |
memv — open-source memory for AI agents that only stores what it failed to predict | 27 | I built an open-source memory system for AI agents with a different approach to knowledge extraction.
The problem: Most memory systems extract every fact from conversations and rely on retrieval to sort out what matters. This leads to noisy knowledge bases full of redundant information.
The approach: memv uses predic... | 2026-02-10T18:31:16 | https://www.reddit.com/r/LocalLLaMA/comments/1r18v9c/memv_opensource_memory_for_ai_agents_that_only/ | brgsk | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18v9c | false | null | t3_1r18v9c | /r/LocalLLaMA/comments/1r18v9c/memv_opensource_memory_for_ai_agents_that_only/ | false | false | self | 27 | null |
Is qwen3 next the real deal? | 1 | Helo safe lamers,
I usually work with claude/copilot in vscode with tools mcp and extensions i built for my workflows, everything ok.
I also use local model up to 16gb mac ram m4… let say qwen2 14b for example or lfm for tooling layers and so.
I am quite happy by do tooling with qwen3:8b and 4b but as far I heard th... | 2026-02-10T18:27:26 | https://www.reddit.com/r/LocalLLaMA/comments/1r18r8c/is_qwen3_next_the_real_deal/ | fab_space | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18r8c | false | null | t3_1r18r8c | /r/LocalLLaMA/comments/1r18r8c/is_qwen3_next_the_real_deal/ | false | false | self | 1 | null |
Is IK-Llama-CPP still worth it for CPU offloading scenarios? | 0 | Using ROCm currently with dual GPUs. 48GB on VRAM, ~40GB of experts offloaded into DDR4.
I haven't looked at ik Llama CPP in a while but I see it referenced less and less around here. Is it still worth trying? It's getting pretty regular commits still I see. | 2026-02-10T18:21:39 | https://www.reddit.com/r/LocalLLaMA/comments/1r18le3/is_ikllamacpp_still_worth_it_for_cpu_offloading/ | ForsookComparison | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18le3 | false | null | t3_1r18le3 | /r/LocalLLaMA/comments/1r18le3/is_ikllamacpp_still_worth_it_for_cpu_offloading/ | false | false | self | 0 | null |
built a self-hosted API proxy that strips PII before prompts reach any LLM - works with Ollama too | 0 | been working on this for a while now. started because I'm in australia and kept running into the same problem with clients - they want to use LLMs but compliance won't let them because prompts might contain tax file numbers, medicare details, patient data, etc.
so I built a proxy that sits between your apps and whatev... | 2026-02-10T18:20:24 | https://www.reddit.com/r/LocalLLaMA/comments/1r18k4m/built_a_selfhosted_api_proxy_that_strips_pii/ | Tradi3 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18k4m | false | null | t3_1r18k4m | /r/LocalLLaMA/comments/1r18k4m/built_a_selfhosted_api_proxy_that_strips_pii/ | false | false | self | 0 | null |
Это полный позор: зачем надо было вводить ограниченную возможность просмотра текста песни в YouTube Music? | 1 | [removed] | 2026-02-10T18:16:18 | https://www.reddit.com/r/LocalLLaMA/comments/1r18fvq/это_полный_позор_зачем_надо_было_вводить/ | TheCoffinOfAndy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18fvq | false | null | t3_1r18fvq | /r/LocalLLaMA/comments/1r18fvq/это_полный_позор_зачем_надо_было_вводить/ | false | false | self | 1 | null |
CPU Usage is diffrent between swepplamabench and lamaserver *IK lamacpp* | 1 | [lamaserver.exe](https://preview.redd.it/9hizcz8vkpig1.png?width=1051&format=png&auto=webp&s=0263720ecf8e38c7a380b8d007f22a27b1c8de22)
[sweeplamabench](https://preview.redd.it/a5eayqr2kpig1.png?width=1117&format=png&auto=webp&s=d409930389cb826a6143a1f8d555d90cffdea1e8)
on ik lamacpp why does lama server use only 40% ... | 2026-02-10T18:15:21 | https://www.reddit.com/r/LocalLLaMA/comments/1r18ewb/cpu_usage_is_diffrent_between_swepplamabench_and/ | Noobysz | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r18ewb | false | null | t3_1r18ewb | /r/LocalLLaMA/comments/1r18ewb/cpu_usage_is_diffrent_between_swepplamabench_and/ | false | false | 1 | null | |
Preprocessing and prompt formatting with multimodal models in llama.cpp | 1 | I have some coding experiences but am still pretty new to AI. So far I managed to set up a few local inferences, but I struggled with understanding the right preprocessing and more important prompt message formatting.
Example: [https://huggingface.co/dam2452/Qwen3-VL-Embedding-8B-GGUF](https://huggingface.co/dam245... | 2026-02-10T18:08:59 | https://www.reddit.com/r/LocalLLaMA/comments/1r188gy/preprocessing_and_prompt_formatting_with/ | AdaObvlada | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r188gy | false | null | t3_1r188gy | /r/LocalLLaMA/comments/1r188gy/preprocessing_and_prompt_formatting_with/ | false | false | self | 1 | {'enabled': False, 'images': [{'id': '6f7BmgVUKhk_KNf-4zCrgQ0KOeumPXS1xPUfzKwGnE8', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/6f7BmgVUKhk_KNf-4zCrgQ0KOeumPXS1xPUfzKwGnE8.png?width=108&crop=smart&auto=webp&s=05400029db7557e770087e39a9114c5672f28c24', 'width': 108}, {'height': 116, 'url': 'h... |
Building a cognitive architecture for emergent AI identity — blank slate to selfhood | 0 | \# intuitive-AI
\*\*What happens when you give an AI the machinery for selfhood but seed it with nothing?\*\*
This project is an experiment in emergent identity. It builds an autonomous agent with layered memory, metacognition, goal formation, creative impulse, and the functional equivalent of an unconscious mind... | 2026-02-10T17:51:11 | https://www.reddit.com/r/LocalLLaMA/comments/1r17qbj/building_a_cognitive_architecture_for_emergent_ai/ | alecprats | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r17qbj | false | null | t3_1r17qbj | /r/LocalLLaMA/comments/1r17qbj/building_a_cognitive_architecture_for_emergent_ai/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'JEoFLjV9kG2DMueLS9yu_EyxIh4_N6bb2EZq8MCq8Qw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/JEoFLjV9kG2DMueLS9yu_EyxIh4_N6bb2EZq8MCq8Qw.png?width=108&crop=smart&auto=webp&s=0306aa8a65ca71f47b2ace0e278c1c440d664bc5', 'width': 108}, {'height': 108, 'url': 'h... |
After 1.5M API keys leaked from OpenClaw, here's how Nucleus MCP prevents sleeper agents | 0 | ```
Watched this video about OpenClaw's security crisis: https://www.youtube.com/watch?v=ceEUO_i7aW4
TLDR: Sleeper agents in top ClawHub skills, Docker escapes, 1.5M API keys leaked from chat logs.
I've been building Nucleus MCP — a local-first MCP server with security built in from day one:
**What it already has (s... | 2026-02-10T17:49:09 | https://www.reddit.com/r/LocalLLaMA/comments/1r17ob6/after_15m_api_keys_leaked_from_openclaw_heres_how/ | NucleusOS | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r17ob6 | false | null | t3_1r17ob6 | /r/LocalLLaMA/comments/1r17ob6/after_15m_api_keys_leaked_from_openclaw_heres_how/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'VtGkJA8oJPl1DNwl7sVUp0xZMInFTbMLgflbpW7oX_M', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/VtGkJA8oJPl1DNwl7sVUp0xZMInFTbMLgflbpW7oX_M.jpeg?width=108&crop=smart&auto=webp&s=93d3c8423c721ed025cfe638476f28e742a90477', 'width': 108}, {'height': 162, 'url': '... |
Is the Nvidia T4 actually viable for 70B (EXL2) daily driving, or is it just pure cope compared to dual 3090s? | 2 | I’ve been trying to find a middle ground for running 70B parameter models without dropping $1.5k on a dual 3090 rig or dealing with the power bill/noise of enterprise used gear (looking at you, P40 screamers).
My local setup (single 3070) is fine for 8B models, but it chokes hard on anything substantial unless I q... | 2026-02-10T17:48:16 | https://www.reddit.com/r/LocalLLaMA/comments/1r17ng4/is_the_nvidia_t4_actually_viable_for_70b_exl2/ | DenisRoger001 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r17ng4 | false | null | t3_1r17ng4 | /r/LocalLLaMA/comments/1r17ng4/is_the_nvidia_t4_actually_viable_for_70b_exl2/ | false | false | self | 2 | null |
The fastest way to run qwen3 localy | 0 | I tryed tò run the following model :
https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8
Using theese software:
Lama.cpp,kobold.cpp, ollama
They are slow
My gpu 2060 6gbvram | 2026-02-10T17:36:45 | https://www.reddit.com/r/LocalLLaMA/comments/1r17bvh/the_fastest_way_to_run_qwen3_localy/ | Quiet_Dasy | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r17bvh | false | null | t3_1r17bvh | /r/LocalLLaMA/comments/1r17bvh/the_fastest_way_to_run_qwen3_localy/ | false | false | self | 0 | {'enabled': False, 'images': [{'id': 'iv8bITQ9TwtfnhpoBEDq1e0iSLbjcAS5l3v64sg728o', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/iv8bITQ9TwtfnhpoBEDq1e0iSLbjcAS5l3v64sg728o.png?width=108&crop=smart&auto=webp&s=9045ab5b815a6daba361a062c147f18afef73b93', 'width': 108}, {'height': 116, 'url': 'h... |
Huawei CSL just quietly dropped its first GGUF models quantized with a SINQ variant | 1 | [removed] | 2026-02-10T17:36:01 | https://www.reddit.com/r/LocalLLaMA/comments/1r17b5o/huawei_csl_just_quietly_dropped_its_first_gguf/ | Only-Care-6333 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r17b5o | false | null | t3_1r17b5o | /r/LocalLLaMA/comments/1r17b5o/huawei_csl_just_quietly_dropped_its_first_gguf/ | false | false | 1 | null | |
Has anyone seen grokking during LLM fine-tuning? What works in practice? | 3 | Hi everyone,
I’ve been reading about the idea of grokking in model training — e.g., a sudden jump in generalization after initial overfitting — and I’m curious how (or whether) this phenomenon applies to fine-tuning LLMs.
A few specific questions:
1. Does grokking actually occur in LLM fine-tuning? Are there publis... | 2026-02-10T17:30:51 | https://www.reddit.com/r/LocalLLaMA/comments/1r1760x/has_anyone_seen_grokking_during_llm_finetuning/ | Fragrant_Presence_98 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r1760x | false | null | t3_1r1760x | /r/LocalLLaMA/comments/1r1760x/has_anyone_seen_grokking_during_llm_finetuning/ | false | false | self | 3 | null |
ktop is a themed terminal system monitor ideal for local LLM setups on Linux (like btop + nvtop) | 97 | I'm working on a hybrid LLM runtime (GPU prefill / CPU inference) and I got tired of switching tabs between nvtop and btop so I built a terminal system monitor that shows both GPUs and CPU (and other good stuff) and also supports themes.
[link to ktop on github](https://github.com/brontoguana/ktop) | 2026-02-10T17:26:48 | mrstoatey | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r171yu | false | null | t3_1r171yu | /r/LocalLLaMA/comments/1r171yu/ktop_is_a_themed_terminal_system_monitor_ideal/ | false | false | 97 | {'enabled': True, 'images': [{'id': 'q3cpicl4cpig1', 'resolutions': [{'height': 53, 'url': 'https://preview.redd.it/q3cpicl4cpig1.png?width=108&crop=smart&auto=webp&s=667f819776edc87d85e1d756ea025dcc4a29fb3d', 'width': 108}, {'height': 107, 'url': 'https://preview.redd.it/q3cpicl4cpig1.png?width=216&crop=smart&auto=web... | ||
Cooling & build advice for H200s | 0 | Hello! I was tasked with building a **bare-metal inference cluster** at work, and I’m trying to avoid any thermal / performance surprises with **2× H200** in a single node.
I’d love feedback from folks who’ve actually run **H100/H200 PCIe** in self-built (non-OEM) boxes:
* How are you cooling them in practice?
* Are ... | 2026-02-10T17:24:00 | https://www.reddit.com/r/LocalLLaMA/comments/1r16z7k/cooling_build_advice_for_h200s/ | Capable-Strategy-656 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16z7k | false | null | t3_1r16z7k | /r/LocalLLaMA/comments/1r16z7k/cooling_build_advice_for_h200s/ | false | false | self | 0 | null |
ktop is a themed terminal system monitor ideal for local LLM setups on Linux (like btop + nvtop) | 1 |
[ktop themed system monitor](https://preview.redd.it/p1w1x8rn5pig1.png?width=3781&format=png&auto=webp&s=52753d7551386f76cda28239f5a3aa950ec71cec)
I'm working on a hybrid LLM runtime (GPU prefill / CPU inference) and I got tired of switching tabs between nvtop and btop so I built a terminal system monitor that shows... | 2026-02-10T17:22:42 | https://www.reddit.com/r/LocalLLaMA/comments/1r16xw8/ktop_is_a_themed_terminal_system_monitor_ideal/ | mrstoatey | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16xw8 | false | null | t3_1r16xw8 | /r/LocalLLaMA/comments/1r16xw8/ktop_is_a_themed_terminal_system_monitor_ideal/ | false | false | 1 | null | |
Hello guys need some suggestions? | 4 | Hello guys
Recently I started working on creating a custom AI assistant using two LLMs, one as a router to call tools or find the intent of questions, and the other LLM as the brain to reason or answer them.
The problem I am facing is that the router is unable to find extra intent for some questio... | 2026-02-10T17:20:54 | https://www.reddit.com/r/LocalLLaMA/comments/1r16w1i/hello_guys_need_some_suggestions/ | MR___Phantom | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16w1i | false | null | t3_1r16w1i | /r/LocalLLaMA/comments/1r16w1i/hello_guys_need_some_suggestions/ | false | false | self | 4 | null |
Shipped a big AgentCrawl update: robots/sitemaps, disk caching, resumable crawls, structured metadata + chunking | 0 | update from my last [post](https://www.reddit.com/r/LocalLLaMA/comments/1qhc1o0/i_built_a_lightweight_typesafe_web_scraper/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button)
Shipped a big AgentCrawl update: robots/sitemaps, disk caching, resumable crawls, structured metadata + c... | 2026-02-10T17:20:08 | https://www.reddit.com/r/LocalLLaMA/comments/1r16v8r/shipped_a_big_agentcrawl_update_robotssitemaps/ | eatsleepliftcode | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16v8r | false | null | t3_1r16v8r | /r/LocalLLaMA/comments/1r16v8r/shipped_a_big_agentcrawl_update_robotssitemaps/ | false | false | self | 0 | null |
Débutant LLM : faire tourner des modèles quantifiés (4 bits) sur RTX 5090 sous Windows | 0 | Bonjour,
Je débute dans l’utilisation des LLM.
Mon entreprise a récemment acquis une RTX 5090 afin de faire tourner un LLM intégré dans un système de RAG.
J’ai réussi à faire fonctionner CUDA 12.8 avec une version nightly de PyTorch, et je parviens à exécuter un modèle 7B avec Transformers en FP16. En revanche, celu... | 2026-02-10T17:20:03 | https://www.reddit.com/r/LocalLLaMA/comments/1r16v5e/débutant_llm_faire_tourner_des_modèles_quantifiés/ | Numerous_Jellyfish56 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16v5e | false | null | t3_1r16v5e | /r/LocalLLaMA/comments/1r16v5e/débutant_llm_faire_tourner_des_modèles_quantifiés/ | false | false | self | 0 | null |
Trouble getting Qwen3-Coder-Next running | 2 | I am having tons of trouble getting a usable speed out of Qwen3-Coder-Next on my local system:
* Intel i7-12700K
* 48GB DDR4-3200
* RTX 5060 Ti 16GB
* RTX 3060 12GB
I came across [this post here](https://www.reddit.com/r/LocalLLaMA/comments/1qz5uww/qwen3_coder_next_as_first_usable_coding_model_60/) claiming to get 30... | 2026-02-10T17:16:00 | https://www.reddit.com/r/LocalLLaMA/comments/1r16r20/trouble_getting_qwen3codernext_running/ | New-Gate7443 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16r20 | false | null | t3_1r16r20 | /r/LocalLLaMA/comments/1r16r20/trouble_getting_qwen3codernext_running/ | false | false | 2 | null | |
Looking for a local model that can handle Shavian. | 2 | I’ve been playing around with Shavian transliteration in LLMs, specifically Gemini flash, which seems to be able to handle and respond perfectly in Shavian if I set up the context correct, but I haven’t found any local model that can do the same.
I really thought this would be basic enough that any model could handle ... | 2026-02-10T16:59:28 | https://www.reddit.com/r/LocalLLaMA/comments/1r16aen/looking_for_a_local_model_that_can_handle_shavian/ | ElementaryZX | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r16aen | false | null | t3_1r16aen | /r/LocalLLaMA/comments/1r16aen/looking_for_a_local_model_that_can_handle_shavian/ | false | false | self | 2 | null |
[ Removed by moderator ] | 1 | [removed] | 2026-02-10T16:55:59 | https://www.reddit.com/r/LocalLLaMA/comments/1r166ye/llada21_vs_qwen330ba3b_892_tps_at_comparable/ | Exact-Literature-395 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r166ye | false | null | t3_1r166ye | /r/LocalLLaMA/comments/1r166ye/llada21_vs_qwen330ba3b_892_tps_at_comparable/ | false | false | null | 1 | null |
Sub-1-Bit LLM Quantization | 57 | Hey everyone, I’ve been interested in extreme compression, and released [NanoQuant](https://arxiv.org/abs/2602.06694), a quantization method that enables sub-1-bit LLMs.
Sub-binary performance was better than 2-bit GPTQ and the extreme memory compression made custom kernels really fast, but the performance wasn't near... | 2026-02-10T16:39:50 | https://www.reddit.com/r/LocalLLaMA/comments/1r15qqc/sub1bit_llm_quantization/ | d77chong | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r15qqc | false | null | t3_1r15qqc | /r/LocalLLaMA/comments/1r15qqc/sub1bit_llm_quantization/ | false | false | self | 57 | null |
What voice quality metrics actually work for conversational TTS? | 3 | I’m researching how teams evaluate **voice quality** in conversational TTS for real agents (naturalness, prosody, consistency, expressiveness).
Curious what works *in practice*:
* Which voice quality metrics do you rely on today (MOS, MUSHRA, Word Error Rate, etc.)?
* Which ones fail to reflect real conversational ex... | 2026-02-10T16:33:20 | https://www.reddit.com/r/LocalLLaMA/comments/1r15kgf/what_voice_quality_metrics_actually_work_for/ | Envelope-Labs | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r15kgf | false | null | t3_1r15kgf | /r/LocalLLaMA/comments/1r15kgf/what_voice_quality_metrics_actually_work_for/ | false | false | self | 3 | null |
Which model Is The fastest for my setup:1650(4gb)? | 0 | 326 MB - model (fp32)
305 MB - model_q4 (4-bit 0 matmul)
177 MB - model_uint8 (8-bit 8 mixed precision)
163 MB - model_fp16 (fp16)
154 MB - model_q4f16 (4-bit 0 matmul & fp16 weights)
114 MB - model_uint8f16 (Mixed precision)
92.4 MB - model_quantized (8-bit)
86 MB - model_q8f16 | 2026-02-10T16:12:04 | Quiet_Dasy | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r14z4i | false | null | t3_1r14z4i | /r/LocalLLaMA/comments/1r14z4i/which_model_is_the_fastest_for_my_setup16504gb/ | false | false | 0 | {'enabled': True, 'images': [{'id': 'NAwBaMPOkXrV9LO1qxjnXZOTAPNQ1Lgpp_7MTtgmliY', 'resolutions': [{'height': 145, 'url': 'https://preview.redd.it/svspx50xyoig1.jpeg?width=108&crop=smart&auto=webp&s=9277e8f05f11e58a7d54969eedf520f02c6e810c', 'width': 108}, {'height': 290, 'url': 'https://preview.redd.it/svspx50xyoig1.j... | ||
Qwen 3 TTS is streaming even working? | 8 | Hey guys,
I'm playing around with Qwen3-TTS for a voice-agent POC and I cant get streaming working.
The docs mention streaming, but I can’t seem to get streaming generation working in practice (even with Claude’s help). What I’m trying to do is have TTS start generating audio as soon as it parses some partial text, ... | 2026-02-10T16:11:54 | https://www.reddit.com/r/LocalLLaMA/comments/1r14yyv/qwen_3_tts_is_streaming_even_working/ | vasa133769 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r14yyv | false | null | t3_1r14yyv | /r/LocalLLaMA/comments/1r14yyv/qwen_3_tts_is_streaming_even_working/ | false | false | self | 8 | null |
Tether: Claude / Codex -> Telegram / Discord / Slack | 0 | With some tasks I felt like i was just reading and clicking 'yes' to permission prompts. I figured I could do that while lunching as well, or from the bathroom. So I built Tether. It has a local-first web UI, but I myself use it through Discord. Has MCP server support too, so Claude can also talk through it directly if... | 2026-02-10T16:05:43 | https://www.reddit.com/r/LocalLLaMA/comments/1r14sw0/tether_claude_codex_telegram_discord_slack/ | wouldacouldashoulda | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r14sw0 | false | null | t3_1r14sw0 | /r/LocalLLaMA/comments/1r14sw0/tether_claude_codex_telegram_discord_slack/ | false | false | self | 0 | null |
How are folks running large dense models on home gear? | 0 | I have a dual RTX 5060 Ti desktop, 32GB VRAM total as my first AI learning box. Later I felt I wanted to run larger models, so I got a NVIDIA Thor Dev kit, and I also played with AI on a 64GB Macbook. In all cases, I find that a 4 bit quantized model with 3B active parameters runs fast so long as it fits in video or un... | 2026-02-10T16:03:10 | https://www.reddit.com/r/LocalLLaMA/comments/1r14qeb/how_are_folks_running_large_dense_models_on_home/ | catplusplusok | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r14qeb | false | null | t3_1r14qeb | /r/LocalLLaMA/comments/1r14qeb/how_are_folks_running_large_dense_models_on_home/ | false | false | self | 0 | null |
How To Run Qwen3-Coder-Next (Q4) on RTX 5060 Ti 16GB + RTX 3060 12GB | 1 | I am having tons of trouble getting a usable speed out of Qwen3-Coder-Next on my local system:
* Intel i7-12700K
* 48GB DDR4-3200
* RTX 5060 Ti 16GB
* RTX 3060 12GB
I came across [this post here](https://www.reddit.com/r/LocalLLaMA/comments/1qz5uww/qwen3_coder_next_as_first_usable_coding_model_60/) claiming to get 30... | 2026-02-10T16:00:17 | https://www.reddit.com/r/LocalLLaMA/comments/1r14ngd/how_to_run_qwen3codernext_q4_on_rtx_5060_ti_16gb/ | New-Gate7443 | self.LocalLLaMA | 1970-01-01T00:00:00 | 0 | {} | 1r14ngd | false | null | t3_1r14ngd | /r/LocalLLaMA/comments/1r14ngd/how_to_run_qwen3codernext_q4_on_rtx_5060_ti_16gb/ | false | false | self | 1 | null |
Train MoE models 12x faster with 30% less memory! (<15GB VRAM) | 409 | Hey [r/LocalLlama]()! We’re excited to introduce \~12x faster Mixture of Experts (MoE) training with **>35% less VRAM** and **\~6x longer context** via our new custom Triton kernels and math optimizations (no accuracy loss). Unsloth repo: [https://github.com/unslothai/unsloth](https://github.com/unslothai/unsloth)
* U... | 2026-02-10T15:54:02 | danielhanchen | i.redd.it | 1970-01-01T00:00:00 | 0 | {} | 1r14h9u | false | null | t3_1r14h9u | /r/LocalLLaMA/comments/1r14h9u/train_moe_models_12x_faster_with_30_less_memory/ | false | false | 409 | {'enabled': True, 'images': [{'id': 'ee2jwnijvoig1', 'resolutions': [{'height': 119, 'url': 'https://preview.redd.it/ee2jwnijvoig1.png?width=108&crop=smart&auto=webp&s=318387fc11fd093e499cdecd44cfcd0b8b9160d4', 'width': 108}, {'height': 238, 'url': 'https://preview.redd.it/ee2jwnijvoig1.png?width=216&crop=smart&auto=we... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.