name
string
body
string
score
int64
controversiality
int64
created
timestamp[us]
author
string
collapsed
bool
edited
timestamp[us]
gilded
int64
id
string
locked
bool
permalink
string
stickied
bool
ups
int64
t1_o7xhxo3
Looks cool!
3
0
2026-02-28T19:35:25
lundrog
false
null
0
o7xhxo3
false
/r/LocalLLaMA/comments/1rhbfya/shunyanet_sentinel_a_selfhosted_rss_aggregator/o7xhxo3/
false
3
t1_o7xhw05
Just don't type "hello"
1
0
2026-02-28T19:35:10
zipzag
false
null
0
o7xhw05
false
/r/LocalLLaMA/comments/1rh43za/qwen_3535ba3b_is_beyond_expectations_its_replaced/o7xhw05/
false
1
t1_o7xhqa5
And it is actually punching above its weight (but not usable for me due to the insane thinking times!, would just tune a bigger model that would take less time I guess!)
5
0
2026-02-28T19:34:20
Potential_Block4598
false
null
0
o7xhqa5
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xhqa5/
false
5
t1_o7xhmz3
No. In fact v4 is coming soon according to Hassabis
1
0
2026-02-28T19:33:51
alexx_kidd
false
null
0
o7xhmz3
false
/r/LocalLLaMA/comments/1rh3thm/rip_gemma_leave_your_memories_here/o7xhmz3/
false
1
t1_o7xhmrk
Have you tried nanbeige? It is a 4B model that thinks A LOT (one question might take 3k tokens of thinking!)
7
0
2026-02-28T19:33:49
Potential_Block4598
false
null
0
o7xhmrk
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xhmrk/
false
7
t1_o7xhlnn
Which is dumb because GPTOSS is also a MoE, you are comparing Apples to Apples already, no formula needed gpt-oss-120B has 5.1B acrmtive parameters on the MoE Layers, and the MoE layers are trained from the ground up in MXFP4 format. That formula is for comparing dense and MoE models, but is kinda outdated because arc...
4
0
2026-02-28T19:33:39
guesdo
false
null
0
o7xhlnn
false
/r/LocalLLaMA/comments/1rh43za/qwen_3535ba3b_is_beyond_expectations_its_replaced/o7xhlnn/
false
4
t1_o7xhkqo
Traces is the fix
1
0
2026-02-28T19:33:32
mikkel1156
false
null
0
o7xhkqo
false
/r/LocalLLaMA/comments/1rgwyqi/agent_debugging_is_a_mess_am_i_the_only_one/o7xhkqo/
false
1
t1_o7xhkm1
ngl, this sub is my favorite place on the internet lately.
8
0
2026-02-28T19:33:31
jeremyckahn
false
null
0
o7xhkm1
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xhkm1/
false
8
t1_o7xhk7t
what model are you using for coding?
1
0
2026-02-28T19:33:27
msbeaute00000001
false
null
0
o7xhk7t
false
/r/LocalLLaMA/comments/1rh46g2/why_some_still_playing_with_old_models_nostalgia/o7xhk7t/
false
1
t1_o7xhixs
I was referring to the overall concept that overthinking things can lead to worse accuracy
1
0
2026-02-28T19:33:15
ThatRandomJew7
false
null
0
o7xhixs
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xhixs/
false
1
t1_o7xhh5l
Qwen is just as good as Gemma in Greek
1
0
2026-02-28T19:32:59
alexx_kidd
false
null
0
o7xhh5l
false
/r/LocalLLaMA/comments/1rh3thm/rip_gemma_leave_your_memories_here/o7xhh5l/
false
1
t1_o7xhecf
I've asked Gemini that question. The model is meant to be different, to my surprise, at least some of the time. I would suggest you pay for one month after planning a systematic way to test both modes, save the output, and get a different cloud model to analyse. You can ask free Claude. Then depending on the outcome yo...
-2
0
2026-02-28T19:32:34
Hector_Rvkp
false
null
0
o7xhecf
false
/r/LocalLLaMA/comments/1rhbbq1/gemini_ultra_vs_pro_actually_different_or_just_a/o7xhecf/
false
-2
t1_o7xhckt
9 what nine ?! Qwen3.5 the big one 122b 27b And the 35b That is not nine is it ?!
1
0
2026-02-28T19:32:19
Potential_Block4598
false
null
0
o7xhckt
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xhckt/
false
1
t1_o7xhbtu
We need wrappers like LM Studio to add the presence penalty that helps mitigate that.
1
0
2026-02-28T19:32:12
Iory1998
false
null
0
o7xhbtu
false
/r/LocalLLaMA/comments/1rhaoty/anyone_noticing_qwen35_27b_getting_stuck_in/o7xhbtu/
false
1
t1_o7xh7h8
Not quite – prefix caching helps when multiple requests share the same prompt prefix (like a system prompt). But in a multi agent chain, each agent’s prompt is different, it includes the previous agent’s output. So there’s no shared prefix to cache between hops. AVP skips that entirely. Instead of pasting text output ...
5
0
2026-02-28T19:31:34
proggmouse
false
null
0
o7xh7h8
false
/r/LocalLLaMA/comments/1rh802w/what_if_llm_agents_passed_kvcache_to_each_other/o7xh7h8/
false
5
t1_o7xh7iz
This is not what the paper states.
0
0
2026-02-28T19:31:34
Thomas-Lore
false
null
0
o7xh7iz
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xh7iz/
false
0
t1_o7xh5b5
> I came across this article on 𝕏 where they used Clawdbot with polymarket to make money. The ratio of red flags to words in this sentence is amazing.
5
0
2026-02-28T19:31:15
TastesLikeOwlbear
false
null
0
o7xh5b5
false
/r/LocalLLaMA/comments/1rhazbc/13_m1_mbp_instead_of_m4_mac_mini/o7xh5b5/
false
5
t1_o7xh2qt
Haven’t tried Claude code but I know from vllm install and trying Ollama there are a lot of stability issues with dependency versions so it may need some time still to be stabilized. Once o got it working was impressive for local
2
0
2026-02-28T19:30:52
2BucChuck
false
null
0
o7xh2qt
false
/r/LocalLLaMA/comments/1rh6455/anybody_able_to_get_qwen3535ba3b_working_with/o7xh2qt/
false
2
t1_o7xgxbs
This is not what the paper states.
12
0
2026-02-28T19:30:05
Thomas-Lore
false
null
0
o7xgxbs
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xgxbs/
false
12
t1_o7xgws6
did anyone try it with the ai hat yet?
1
0
2026-02-28T19:30:01
overflow74
false
null
0
o7xgws6
false
/r/LocalLLaMA/comments/1rg87bj/qwen3535ba3b_running_on_a_raspberry_pi_5_16gb_and/o7xgws6/
false
1
t1_o7xgwmm
They are not on the same level. Qwen3.5-35B-A3B-GGUF Q5_K_M is 26.2GB.
0
0
2026-02-28T19:29:59
moahmo88
false
null
0
o7xgwmm
false
/r/LocalLLaMA/comments/1rh0msv/qwen3535ba3b_q5_k_mbest_model_for_nvidia_16gb_gpus/o7xgwmm/
false
0
t1_o7xgugq
At only 50 tokens? I doubt it.
1
0
2026-02-28T19:29:40
Thomas-Lore
false
null
0
o7xgugq
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xgugq/
false
1
t1_o7xgtzg
I have a disk of x teras, which models should I download? They are recurrent but at the same time very useful. The problem comes later with a new model that I eliminate? If I delete the version before this, what am I losing?
1
0
2026-02-28T19:29:36
Macestudios32
false
null
0
o7xgtzg
false
/r/LocalLLaMA/comments/1rguty0/get_your_local_models_in_order_anthropic_just_got/o7xgtzg/
false
1
t1_o7xgsi6
Who gives a shit if it was prompt engineered or not? Automation doesn't negate anything. You're a dipshit and will be left in the past. Have fun with that
1
0
2026-02-28T19:29:23
TheBurkMeister
false
null
0
o7xgsi6
false
/r/LocalLLaMA/comments/1r9gve8/i_feel_left_behind_what_is_special_about_openclaw/o7xgsi6/
false
1
t1_o7xgsc3
? Many MLX quants are already available. I'm running it using oMLX just fine
7
0
2026-02-28T19:29:22
dwkdnvr
false
null
0
o7xgsc3
false
/r/LocalLLaMA/comments/1rh9k63/qwen35_35ba3b_replaced_my_2model_agentic_setup_on/o7xgsc3/
false
7
t1_o7xgp6t
Bots are here to stay
9
0
2026-02-28T19:28:55
DinoAmino
false
null
0
o7xgp6t
false
/r/LocalLLaMA/comments/1rhb1xb/fine_tuning_on_proprietary_data_is_way_harder_to/o7xgp6t/
false
9
t1_o7xglrp
They can think what ever they want. And not every thought needs to be hurled at someone else to make them feel like shit
1
0
2026-02-28T19:28:25
TheBurkMeister
false
null
0
o7xglrp
false
/r/LocalLLaMA/comments/1r9gve8/i_feel_left_behind_what_is_special_about_openclaw/o7xglrp/
false
1
t1_o7xglls
Local models are the future. The consolidation is happening soon for the megacorps. Many will be burned.
1
0
2026-02-28T19:28:24
Taki_Minase
false
null
0
o7xglls
false
/r/LocalLLaMA/comments/1rguty0/get_your_local_models_in_order_anthropic_just_got/o7xglls/
false
1
t1_o7xgkdi
that’s true. apologies for the inconvenience. is there a way you can run the following command on your terminal. it’s a HF command : # delete any partial blob and force fresh download rm -rf ~/.cache/huggingface/hub/models--srswti--bodega-raptor-0.9b/blobs/* # force download with explicit file huggingface-cli downlo...
1
0
2026-02-28T19:28:14
EmbarrassedAsk2887
false
null
0
o7xgkdi
false
/r/LocalLLaMA/comments/1rgkzlo/realtime_speech_to_speech_engine_runs_fully_local/o7xgkdi/
false
1
t1_o7xghy7
Precompiled Llama.cpp is great as it comes with it's own webui. './llama-server -m /path/to/model' open up browser, input 'localhost:8080' and you're golden.
1
0
2026-02-28T19:27:53
hackiv
false
null
0
o7xghy7
false
/r/LocalLLaMA/comments/1rh1q8j/i919400f_rtx_4070_super_12gb_32gb_ddr5_ram/o7xghy7/
false
1
t1_o7xgcsz
Still using it for conversation loops
1
0
2026-02-28T19:27:09
mikkel1156
false
null
0
o7xgcsz
false
/r/LocalLLaMA/comments/1rh3thm/rip_gemma_leave_your_memories_here/o7xgcsz/
false
1
t1_o7xg9cn
yeah, just seems like extra steps and more links in the 'ol security chain.
2
0
2026-02-28T19:26:38
Complainer_Official
false
null
0
o7xg9cn
false
/r/LocalLLaMA/comments/1rhb99s/agents_are_here_but_ecommerce_stuck_in_the_past/o7xg9cn/
false
2
t1_o7xg8fk
This is not what the paper proposes. They run several branches but only allow some of them to finish thinking, the rest is thrown out. It speeds up the multi-reasoning approach that Deep Thinking version of Gemini or GPT 5 Pro models use.
1
0
2026-02-28T19:26:30
Thomas-Lore
false
null
0
o7xg8fk
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xg8fk/
false
1
t1_o7xg7vf
Yeah Gemini 3 loves to reason in comments above code it writes, haha.
13
0
2026-02-28T19:26:25
dsanft
false
null
0
o7xg7vf
false
/r/LocalLLaMA/comments/1rh5luv/qwen35_35ba3b_evaded_the_zeroreasoning_budget_by/o7xg7vf/
false
13
t1_o7xg6w5
Even at Q3 quants it beats it for me...
1
0
2026-02-28T19:26:17
Significant_Fig_7581
false
null
0
o7xg6w5
false
/r/LocalLLaMA/comments/1rgoygs/does_qwen35_35b_outperform_qwen3_coder_next_80b/o7xg6w5/
false
1
t1_o7xg62k
Windows version? And llama.cpp support?
3
0
2026-02-28T19:26:09
pmttyji
false
null
0
o7xg62k
false
/r/LocalLLaMA/comments/1rhalir/local_llms_are_slow_i_have_too_many_things_to_try/o7xg62k/
false
3
t1_o7xg4ay
I think it’s VS only?
1
0
2026-02-28T19:25:54
BahnMe
false
null
0
o7xg4ay
false
/r/LocalLLaMA/comments/1rgtxry/is_qwen35_a_coding_game_changer_for_anyone_else/o7xg4ay/
false
1
t1_o7xg3vc
TLDR; You need to set your CTX-Size to 2048, to have true Apples to Apples. 2048 is the base height to ride the PPL/KLD ride when comparing models. I did not see this before, but in your information above you have your context set for PPL at 512. This is not apples to apples. I recommend reading up on how Turbo...
2
0
2026-02-28T19:25:50
Phaelon74
false
null
0
o7xg3vc
false
/r/LocalLLaMA/comments/1rgel19/new_qwen3535ba3b_unsloth_dynamic_ggufs_benchmarks/o7xg3vc/
false
2
t1_o7xg0k6
You forgot to specify your parameters, your model quant, your KV cache quant. Please give more context.
1
0
2026-02-28T19:25:21
PhilippeEiffel
false
null
0
o7xg0k6
false
/r/LocalLLaMA/comments/1rhaoty/anyone_noticing_qwen35_27b_getting_stuck_in/o7xg0k6/
false
1
t1_o7xfyhj
So this means that this is useless if I'm using an inference engine that has prefix caching? I feel like all of them do nowadays.
8
0
2026-02-28T19:25:03
No-Refrigerator-1672
false
null
0
o7xfyhj
false
/r/LocalLLaMA/comments/1rh802w/what_if_llm_agents_passed_kvcache_to_each_other/o7xfyhj/
false
8
t1_o7xfum9
What’s a good one to start with?
1
0
2026-02-28T19:24:30
BahnMe
false
null
0
o7xfum9
false
/r/LocalLLaMA/comments/1rgtxry/is_qwen35_a_coding_game_changer_for_anyone_else/o7xfum9/
false
1
t1_o7xfsgz
This is in llama serve? How is the command will it run on gpu and as well cpu if low vram, just a look on the command.
1
0
2026-02-28T19:24:11
abubakkar_s
false
null
0
o7xfsgz
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xfsgz/
false
1
t1_o7xfrhg
I don't think anything on the iPhone is private so what's the use case for the local LLM?
2
0
2026-02-28T19:24:02
jacek2023
false
null
0
o7xfrhg
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xfrhg/
false
2
t1_o7xfodx
Oh I actually like the look of this. I've implemented half of this in slightly hacky ways for my own use but this looks much cleaner. I'll give it a good shot
2
0
2026-02-28T19:23:35
Front_Eagle739
false
null
0
o7xfodx
false
/r/LocalLLaMA/comments/1rhalir/local_llms_are_slow_i_have_too_many_things_to_try/o7xfodx/
false
2
t1_o7xfnzi
isn't it all just Gemini 3.1 Pro?
4
0
2026-02-28T19:23:31
atape_1
false
null
0
o7xfnzi
false
/r/LocalLLaMA/comments/1rhbbq1/gemini_ultra_vs_pro_actually_different_or_just_a/o7xfnzi/
false
4
t1_o7xfnd8
PayPal = your real credentials stored, merchant has to integrate. PayClaw = fresh card per task, anywhere Visa works, expires in 15 min. Closer to Ramp for agents than PayPal.
1
0
2026-02-28T19:23:26
Opposite-Exam3541
false
null
0
o7xfnd8
false
/r/LocalLLaMA/comments/1rhb99s/agents_are_here_but_ecommerce_stuck_in_the_past/o7xfnd8/
false
1
t1_o7xfja9
vLLM has some interesting capabilities: \- robust when concurrency is required \- ability to change gpt-oss reasoning level on the fly (llama.cpp need restart) \- recommended default parameters are in the model's files I probably forgot some additional advantages.
1
0
2026-02-28T19:22:51
PhilippeEiffel
false
null
0
o7xfja9
false
/r/LocalLLaMA/comments/1r7l7q5/the_strix_halo_feels_like_an_amazing_super_power/o7xfja9/
false
1
t1_o7xfhdx
My school give free access to the HPU which contains many 3090s, H200s, RTX 6000, A90s ect. Its been fun
4
0
2026-02-28T19:22:34
jovn1234567890
false
null
0
o7xfhdx
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xfhdx/
false
4
t1_o7xfegu
Maybe frees up gpus at claude, more free tokens coming!
1
0
2026-02-28T19:22:09
Comfortable_Camp9744
false
null
0
o7xfegu
false
/r/LocalLLaMA/comments/1rguty0/get_your_local_models_in_order_anthropic_just_got/o7xfegu/
false
1
t1_o7xfclv
The Chinese are laughing.
1
0
2026-02-28T19:21:53
Own_Respond_9189
false
null
0
o7xfclv
false
/r/LocalLLaMA/comments/1rgn4ki/president_trump_orders_all_federal_agencies_in/o7xfclv/
false
1
t1_o7xfb55
Smart.
1
0
2026-02-28T19:21:41
Pantoffel86
false
null
0
o7xfb55
false
/r/LocalLLaMA/comments/1rh5luv/qwen35_35ba3b_evaded_the_zeroreasoning_budget_by/o7xfb55/
false
1
t1_o7xfar2
The only AI sub not on my mute list! Love it here!
32
0
2026-02-28T19:21:38
AbheekG
false
null
0
o7xfar2
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xfar2/
false
32
t1_o7xf3v6
Hahahaha I laughed a lot harder than expected.
2
0
2026-02-28T19:20:37
Hector_Rvkp
false
null
0
o7xf3v6
false
/r/LocalLLaMA/comments/1rhazbc/13_m1_mbp_instead_of_m4_mac_mini/o7xf3v6/
false
2
t1_o7xf2y7
so... its paypal?
1
0
2026-02-28T19:20:29
Complainer_Official
false
null
0
o7xf2y7
false
/r/LocalLLaMA/comments/1rhb99s/agents_are_here_but_ecommerce_stuck_in_the_past/o7xf2y7/
false
1
t1_o7xf2j1
If it fits an iphone it will be an instant favorite
1
0
2026-02-28T19:20:26
Traditional-Card6096
false
null
0
o7xf2j1
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xf2j1/
false
1
t1_o7xex8p
Yeah, the market is not so welcoming for now, so i decided to be loyal at work now :D
12
0
2026-02-28T19:19:40
bobaburger
false
null
0
o7xex8p
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xex8p/
false
12
t1_o7xew3r
The problem in my mind is you need to have a good underlying simulation. If humans can have fun doing the activities you will learn more from the models.
1
0
2026-02-28T19:19:31
vox-deorum
false
null
0
o7xew3r
false
/r/LocalLLaMA/comments/1rf7wy0/rant_post_genuinely_losing_my_mind_over_a_llm/o7xew3r/
false
1
t1_o7xerws
I use it for a baseline for 2 of my Apps. (They are MacOS only - sorry!) [https://github.com/scouzi1966/maclocal-api](https://github.com/scouzi1966/maclocal-api) (OpenSource) [https://github.com/scouzi1966/vesta-mac-dist](https://github.com/scouzi1966/vesta-mac-dist) (Closed source for now -- want to shape it). Th...
1
0
2026-02-28T19:18:55
scousi
false
null
0
o7xerws
false
/r/LocalLLaMA/comments/1rh1haa/benchmarks_report_optimized_cosmosreason2_qwen3vl/o7xerws/
false
1
t1_o7xepbk
I built one for them to play Civilization. Immense fun. Hopefully to be able to share the detailed data with you soon..
1
0
2026-02-28T19:18:33
vox-deorum
false
null
0
o7xepbk
false
/r/LocalLLaMA/comments/1rf7wy0/rant_post_genuinely_losing_my_mind_over_a_llm/o7xepbk/
false
1
t1_o7xep7f
my PC is slow for local LLMs so I'd kick off a task and just... wait - so im not saying local llms are slow in general sorry
1
0
2026-02-28T19:18:32
BadBoy17Ge
false
null
0
o7xep7f
false
/r/LocalLLaMA/comments/1rhalir/local_llms_are_slow_i_have_too_many_things_to_try/o7xep7f/
false
1
t1_o7xen7r
FWIW LMCache solves a different problem. It caches KV for previously seen text so you don’t re-prefill the same prompt across requests. AVP transfers KV-cache between agents with different prompts as a communication channel. One is “I’ve seen this text before, skip prefill.” The other is “here’s my reasoning, don’t ma...
7
0
2026-02-28T19:18:15
proggmouse
false
null
0
o7xen7r
false
/r/LocalLLaMA/comments/1rh802w/what_if_llm_agents_passed_kvcache_to_each_other/o7xen7r/
false
7
t1_o7xec98
If anything the price of non flagship cards will go up due to increased on premises llm demand lol
3
0
2026-02-28T19:16:42
gradient8
false
null
0
o7xec98
false
/r/LocalLLaMA/comments/1rh095c/deepseek_v4_will_be_released_next_week_and_will/o7xec98/
false
3
t1_o7xe56y
Haha
0
0
2026-02-28T19:15:41
Iory1998
false
null
0
o7xe56y
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xe56y/
false
0
t1_o7xe3j0
I thought that's how tech worked, myself 😂
1
0
2026-02-28T19:15:27
ctanna5
false
null
0
o7xe3j0
false
/r/LocalLLaMA/comments/1rcpmwn/anthropic_weve_identified_industrialscale/o7xe3j0/
false
1
t1_o7xe3dm
4B is good for my purposes actually. I'm writing my own inferencing engine and small models are great to test new architectures with.
1
0
2026-02-28T19:15:25
dsanft
false
null
0
o7xe3dm
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xe3dm/
false
1
t1_o7xdx37
backup: https://files.catbox.moe/dhumvv.png
1
0
2026-02-28T19:14:30
MelodicRecognition7
false
null
0
o7xdx37
false
/r/LocalLLaMA/comments/1rh4rsf/i_compiled_every_confirmed_rubin_vs_blackwell/o7xdx37/
false
1
t1_o7xdun0
My surname is Stalker, and LLMs don't like it because the common noun appears far more often in their training data.
1
0
2026-02-28T19:14:09
ross_st
false
null
0
o7xdun0
false
/r/LocalLLaMA/comments/1rfg3kx/american_closed_models_vs_chinese_open_models_is/o7xdun0/
false
1
t1_o7xdtoy
There is/was a NY court order that prohibited OpenAI from actually deleting the chats that the users "deleted". OpenAI may actually want to be ethical but in the end the US government and US courts can just take the data. And that will cause massive issues in the EU where companies actually have to follow the law.
3
0
2026-02-28T19:14:01
dingo_xd
false
null
0
o7xdtoy
false
/r/LocalLLaMA/comments/1rh43za/qwen_3535ba3b_is_beyond_expectations_its_replaced/o7xdtoy/
false
3
t1_o7xdrc5
Consider big PC as your next computer because you can always replace the future GPU with another future GPU
1
0
2026-02-28T19:13:41
jacek2023
false
null
0
o7xdrc5
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xdrc5/
false
1
t1_o7xdoh3
Never tried it on Windows unfortunately. First try to install llama-server Ask Gemini, it's been really helpful with all kind of installation advice and tuning adive, it's very good at that: [https://gemini.google.com/app](https://gemini.google.com/app)
2
0
2026-02-28T19:13:17
AppealSame4367
false
null
0
o7xdoh3
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xdoh3/
false
2
t1_o7xdljz
Qwen3.5 27B runs ok on 16VRAM at iQ3XS. It can be fully offloaded to GPU and its fast and vison does not cause OOM. But don't expect huge content length. For 35B MOE, it's too much for 16GB🫠 Perhaps IQ2 may fit...not with only 3B active params at Q2... My PC is still DDR4, so medium size MOE models are generally slow ...
2
0
2026-02-28T19:12:52
FancyImagination880
false
null
0
o7xdljz
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xdljz/
false
2
t1_o7xdjab
Sure, but point is to squeeze the hardware that somebody already have. In future we'll have more ad-hoc hardware 
1
0
2026-02-28T19:12:32
Deep_Traffic_7873
false
null
0
o7xdjab
false
/r/LocalLLaMA/comments/1rgixk7/accuracy_vs_speed_my_top_5/o7xdjab/
false
1
t1_o7xdj02
To me it's quite the opposite haha it's funny how personal experience differs from person to person. I have used AesSedai's Minimax 2.5 Q5 to perform incremental tasks in an existing project, and it has been great
6
0
2026-02-28T19:12:29
Septerium
false
null
0
o7xdj02
false
/r/LocalLLaMA/comments/1rgokw1/a_monthly_update_to_my_where_are_openweight/o7xdj02/
false
6
t1_o7xdgb7
look up "GRPO done right"
0
0
2026-02-28T19:12:06
Necessary-Wasabi-619
false
null
0
o7xdgb7
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o7xdgb7/
false
0
t1_o7xdge2
Often, the only way to get that raise is to move firms.
25
0
2026-02-28T19:12:06
Veastli
false
null
0
o7xdge2
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xdge2/
false
25
t1_o7xdfxh
What's the news? Is Gemma officially dead ?
1
0
2026-02-28T19:12:02
tomakorea
false
null
0
o7xdfxh
false
/r/LocalLLaMA/comments/1rh3thm/rip_gemma_leave_your_memories_here/o7xdfxh/
false
1
t1_o7xdel4
It's a laptop, i would have to buy a new one. And if I would do that, I'd buy an HP Elite Book with 128GB unified ram.. Maybe in summer
1
0
2026-02-28T19:11:51
AppealSame4367
false
null
0
o7xdel4
false
/r/LocalLLaMA/comments/1rgzul5/are_you_ready_for_small_qwens/o7xdel4/
false
1
t1_o7xdb0z
I noticed it too - in both qwen3.5-27b and qwen3.5-35b-a3b. It's pretty rare - happened only a few times during my few hundred inferences. I used recommended sampling parameters for precise tasks (temp 0.6, top\_p 0.95, top\_k 20) and official OpenRouter Alibaba endpoint.
1
0
2026-02-28T19:11:21
fairydreaming
false
null
0
o7xdb0z
false
/r/LocalLLaMA/comments/1rhaoty/anyone_noticing_qwen35_27b_getting_stuck_in/o7xdb0z/
false
1
t1_o7xd0tc
the thinking disabled tip is criminally underrated in this post thinking mode is a trap for agentic tasks — you're paying 2-3x latency for marginal gains on steps where the model already knows what to do. the planning overhead kills you in multi-step loops also ditching 2 specialized models removes all the routin...
23
0
2026-02-28T19:09:54
salmenus
false
null
0
o7xd0tc
false
/r/LocalLLaMA/comments/1rh9k63/qwen35_35ba3b_replaced_my_2model_agentic_setup_on/o7xd0tc/
false
23
t1_o7xd0bk
The cleaner path I've seen work is front-loading the data classification conversation before anyone writes a line of code. Get legal to define the data sensitivity tiers and what each tier allows (local inference only, private cloud, managed API with DPA, etc.) in week one. Then your architecture choices fall out of th...
1
0
2026-02-28T19:09:50
ruibranco
false
null
0
o7xd0bk
false
/r/LocalLLaMA/comments/1rhb1xb/fine_tuning_on_proprietary_data_is_way_harder_to/o7xd0bk/
false
1
t1_o7xczxl
thats because google ran trillions of tokens through it, Ive asked gemma 3 27b about a ps1 game called eternal eyes, and it made comments that directly linked to that game, no model at that size can pull as much obscure lore as googles models, because they just dont throw as many tokens at the training as google does
2
0
2026-02-28T19:09:47
Background-Ad-5398
false
null
0
o7xczxl
false
/r/LocalLLaMA/comments/1rh3thm/rip_gemma_leave_your_memories_here/o7xczxl/
false
2
t1_o7xczzu
Is the account selling a course for a crazy price? Then it's a scam
1
0
2026-02-28T19:09:47
Realistic_Muscles
false
null
0
o7xczzu
false
/r/LocalLLaMA/comments/1rhazbc/13_m1_mbp_instead_of_m4_mac_mini/o7xczzu/
false
1
t1_o7xcye6
Disagree and that's fine
1
0
2026-02-28T19:09:33
ForsookComparison
false
null
0
o7xcye6
false
/r/LocalLLaMA/comments/1rgokw1/a_monthly_update_to_my_where_are_openweight/o7xcye6/
false
1
t1_o7xcwp6
sustained performance is lower on a macbook, its for mobility. Also people prefer mini because they already have monitors for their computer setup and mini has less footprint to sit on the desk. Also M4 is just a much higher performing chip? Btw, Clawdbot is not a local LLM, its an API interface basically, so you're ...
1
0
2026-02-28T19:09:18
Amazing_Trace
false
null
0
o7xcwp6
false
/r/LocalLLaMA/comments/1rhazbc/13_m1_mbp_instead_of_m4_mac_mini/o7xcwp6/
false
1
t1_o7xckw6
the article is using an advanced technique known as "lying"
8
0
2026-02-28T19:07:34
HopePupal
false
null
0
o7xckw6
false
/r/LocalLLaMA/comments/1rhazbc/13_m1_mbp_instead_of_m4_mac_mini/o7xckw6/
false
8
t1_o7xckh2
Super will be sort of OP size range. 100B but trained in NVFP4. Ultra is supposed to be 500B.
2
0
2026-02-28T19:07:31
Middle_Bullfrog_6173
false
null
0
o7xckh2
false
/r/LocalLLaMA/comments/1rh9ygz/is_anyone_else_waiting_for_a_6070b_moe_with_810b/o7xckh2/
false
2
t1_o7xciej
have you tried more modern ocr? PaddleVL-OCR? etc?
1
0
2026-02-28T19:07:13
Budget-Juggernaut-68
false
null
0
o7xciej
false
/r/LocalLLaMA/comments/1rh3xey/seeking_help_improving_ocr_in_my_rag_pipeline/o7xciej/
false
1
t1_o7xcglj
Way Worse. Here's KLD on Llama3.1-8B-Instruct. NVFP4 is W4A4 or Weights 4bit Activations 4bit. INT4 for LLM\_Compressor is W4A16 and outperforms NVFP4 by a MILE. TLDR; an NVFP4 HAS to be QAD'd (Quantization Aware Distillation) to do what Nvidia says it can do, which is \~1% of FP8 Quality. NVFP4 and INT4 on Blackw...
3
0
2026-02-28T19:06:58
Phaelon74
false
null
0
o7xcglj
false
/r/LocalLLaMA/comments/1rgel19/new_qwen3535ba3b_unsloth_dynamic_ggufs_benchmarks/o7xcglj/
false
3
t1_o7xcemx
The default llama.cpp batch settings (b=2048, ub=512) are good for TG but pretty bad for PP speeds. The "fit-nobatch" (which means using --fit and not touching the batch setting defaults) may be good if you are only concerned about TG speeds. Personally PP is more important to me so I set ub=2048 even if it hurts TG a ...
1
0
2026-02-28T19:06:42
OsmanthusBloom
false
null
0
o7xcemx
false
/r/LocalLLaMA/comments/1rh9983/qwen_35b_a3b_aessedai_finetune_on_8gb_vram_and/o7xcemx/
false
1
t1_o7xcc7w
16GB but it's HBM, so it has more memory bandwidth than a 3080.
10
0
2026-02-28T19:06:21
FullstackSensei
false
null
0
o7xcc7w
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xcc7w/
false
10
t1_o7xcbfu
The others are right about 64GB being tight... You'll be comfortable with 7B/13B but anything above 30B Q4 will be slow or won't fit cleanly once the OS takes its cut. On the M1 Ultra at £2500 it's a capable machine but you're paying for 2021 silicon. For the same money (or a bit more) the M4 Mac Studio 64GB gets you r...
1
0
2026-02-28T19:06:15
KneeTop2597
false
null
0
o7xcbfu
false
/r/LocalLLaMA/comments/1rha4g1/advice_on_hardware_purchase_and_selling_old/o7xcbfu/
false
1
t1_o7xc5c2
Obligatory [https://xkcd.com/505/](https://xkcd.com/505/)
20
0
2026-02-28T19:05:23
jacobpederson
false
null
0
o7xc5c2
false
/r/LocalLLaMA/comments/1rh9u4r/this_sub_is_incredible/o7xc5c2/
false
20
t1_o7xc4yo
Yup just needa make sure experts are co-located with the compute that routes to them... It's in an awkward zone of needing 2-3 GPUs. Lots of teams jump up to 8...
1
0
2026-02-28T19:05:20
paulahjort
false
null
0
o7xc4yo
false
/r/LocalLLaMA/comments/1rh9ygz/is_anyone_else_waiting_for_a_6070b_moe_with_810b/o7xc4yo/
false
1
t1_o7xbvu5
FTW = LMCache + vLLM https://docs.lmcache.ai/index.html
3
0
2026-02-28T19:04:04
DinoAmino
false
null
0
o7xbvu5
false
/r/LocalLLaMA/comments/1rh802w/what_if_llm_agents_passed_kvcache_to_each_other/o7xbvu5/
false
3
t1_o7xbudn
the biggest win is definitely how well it handles those custom mcps compared to older open source models.. getting it to trigger browser scripts to pull live data instead of just blindly hallucinating an answer makes it actually usable for complex real world workflows..
3
0
2026-02-28T19:03:51
Olivia_Davis_09
false
null
0
o7xbudn
false
/r/LocalLLaMA/comments/1rh43za/qwen_3535ba3b_is_beyond_expectations_its_replaced/o7xbudn/
false
3
t1_o7xbnjh
I will be messaging you in 7 days on [**2026-03-07 19:01:59 UTC**](http://www.wolframalpha.com/input/?i=2026-03-07%2019:01:59%20UTC%20To%20Local%20Time) to remind you of [**this link**](https://www.reddit.com/r/LocalLLaMA/comments/1rh095c/deepseek_v4_will_be_released_next_week_and_will/o7xbgye/?context=3) [**CLICK THI...
1
0
2026-02-28T19:02:55
RemindMeBot
false
null
0
o7xbnjh
false
/r/LocalLLaMA/comments/1rh095c/deepseek_v4_will_be_released_next_week_and_will/o7xbnjh/
false
1
t1_o7xbkaa
i love the attitude but that's not how the corporate world works. they'll accept massive risk in order to save money in the short term, provided there's a legal framework for blaming someone else for hallucinations and data breaches. a previous employer was happily shoveling hundreds of gigabytes of customer images and...
14
0
2026-02-28T19:02:28
HopePupal
false
null
0
o7xbkaa
false
/r/LocalLLaMA/comments/1rh43za/qwen_3535ba3b_is_beyond_expectations_its_replaced/o7xbkaa/
false
14
t1_o7xbi1g
You can allocate as much memory to the GPU as you want with the Strix Halo if you run Linux. Look up 'Strix Halo toolboxes' for the exact instructions. If you buy the motherboard from framework, you can also mount it in a standard case and use an adapter to run your 3090 with it. This is a bit more complex, but, once...
1
0
2026-02-28T19:02:08
jreddit6969
false
null
0
o7xbi1g
false
/r/LocalLLaMA/comments/1rha4g1/advice_on_hardware_purchase_and_selling_old/o7xbi1g/
false
1
t1_o7xbgye
!RemindMe 7 days
1
0
2026-02-28T19:01:59
Stahlboden
false
null
0
o7xbgye
false
/r/LocalLLaMA/comments/1rh095c/deepseek_v4_will_be_released_next_week_and_will/o7xbgye/
false
1
t1_o7xbfwa
The P40s are solid but yeah, 24GB per card starts to hurt once you're running anything above 30B Q4. Multi-GPU offloading helps but the TPS hit is real. If you're thinking about what to move to next, the calculus has shifted a lot in the past year. The Mac Studio M4 Max (64GB unified) hits \~500 GB/s bandwidth and hand...
1
0
2026-02-28T19:01:50
KneeTop2597
false
null
0
o7xbfwa
false
/r/LocalLLaMA/comments/1rha4g1/advice_on_hardware_purchase_and_selling_old/o7xbfwa/
false
1