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_o88en9f
remember, those are MoE. both of those are A3B so they only activate 3B parameters. they should outperform a 3B dense model but they won't be as good as 30B (and esp. 80B) dense, so it makes sense a 9B dense outperforms them. still impressive performance though, for sure.
14
0
2026-03-02T14:24:28
z_latent
false
null
0
o88en9f
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o88en9f/
false
14
t1_o88eh4o
Yep. Auto translated everything.
1
0
2026-03-02T14:23:31
JTN02
false
null
0
o88eh4o
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88eh4o/
false
1
t1_o88egie
Genuine question. What the purpose of such a small model, especially quants like Q2. I can only assume it's for basic commands around home assistant perhaps?
2
0
2026-03-02T14:23:25
sagiroth
false
null
0
o88egie
false
/r/LocalLLaMA/comments/1ritlux/qwen3508bgguf_is_here/o88egie/
false
2
t1_o88e8uh
He did lol
7
0
2026-03-02T14:22:13
Savantskie1
false
null
0
o88e8uh
false
/r/LocalLLaMA/comments/1rit2fy/reverted_from_qwen35_27b_back_to_qwen3_8b/o88e8uh/
false
7
t1_o88e84x
It's your setup, I'm running ROCM 6.10.5 with no issues.
2
0
2026-03-02T14:22:07
MotokoAGI
false
null
0
o88e84x
false
/r/LocalLLaMA/comments/1rikb4w/qwen_35_amd_mi50_32gb_benchmarks/o88e84x/
false
2
t1_o88dsan
At 16gb you could use 30/35b with layer offloading and still get good tok rates. Unless you really need a denser model, I’d probably recommend doing that.
7
0
2026-03-02T14:19:40
3spky5u-oss
false
null
0
o88dsan
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88dsan/
false
7
t1_o88drub
Who's gonna address the elephant in the room?
7
0
2026-03-02T14:19:36
jax_cooper
false
null
0
o88drub
false
/r/LocalLLaMA/comments/1ritr5v/oss120b_beats_all_open_models_but_one_in_new/o88drub/
false
7
t1_o88dkay
Neat idea! Makes me wonder how a game such as Uplink would be like with a LLM involved.
1
0
2026-03-02T14:18:27
Toooooool
false
null
0
o88dkay
false
/r/LocalLLaMA/comments/1riliyt/i_made_a_free_local_ai_roleplay_horror_game/o88dkay/
false
1
t1_o88djsz
According to my benckmarks, there is no improvement related to latest firmware. Using vulkan, I have higher PP and lower tg. I have "-fa on" flag. firmware 20251111 Kernel 6.18.12 llama.cpp b8146 | model | test | t/s | peak t/s | |:------------------|----------------:|----...
1
0
2026-03-02T14:18:23
PhilippeEiffel
false
null
0
o88djsz
false
/r/LocalLLaMA/comments/1ri6yhb/the_last_amd_gpu_firmware_update_together_with/o88djsz/
false
1
t1_o88de9a
are you also on LM Studio0.4.6 (Build 1)? It's been working for me like 90-95% of the time.
1
0
2026-03-02T14:17:31
lolwutdo
false
null
0
o88de9a
false
/r/LocalLLaMA/comments/1ricz8u/notice_qwen_35_reprocessing_the_prompt_every_time/o88de9a/
false
1
t1_o88dcps
it shows glm5 beating gpt-oss-120b. https://preview.redd.it/68i6ny1n4nmg1.png?width=1160&format=png&auto=webp&s=4bbb1224f1d312bd9b13e29481d182839b08550f
3
0
2026-03-02T14:17:17
MotokoAGI
false
null
0
o88dcps
false
/r/LocalLLaMA/comments/1ritr5v/oss120b_beats_all_open_models_but_one_in_new/o88dcps/
false
3
t1_o88d91o
"I spent months believing a flawed Gödel-based argument against AGI because Claude agreed with me." is actually the name of my new math rock band.
1
0
2026-03-02T14:16:43
LoveMind_AI
false
null
0
o88d91o
false
/r/LocalLLaMA/comments/1riovvx/how_not_to_go_insane_talking_with_llms/o88d91o/
false
1
t1_o88d7oy
use llamac++
7
0
2026-03-02T14:16:31
Odd-Ordinary-5922
false
null
0
o88d7oy
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88d7oy/
false
7
t1_o88d6ev
I’m suspicious that new models has some benchmark data (in training data ) probably not deliberatly tho
8
0
2026-03-02T14:16:19
Various-Inside-4064
false
null
0
o88d6ev
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o88d6ev/
false
8
t1_o88d5fx
GLM 4.6V Flash was 9b
3
0
2026-03-02T14:16:10
thejacer
false
null
0
o88d5fx
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88d5fx/
false
3
t1_o88d4tp
just tried out 9b gguf on macbook air 16gb m3, 1sec ttft and 12t/s, i am really impressed and my laptop warmer but i dont care this is nuts. had it make a cheatsheet of top 10 rxjs commands, it popped code examples out and just, wow
1
0
2026-03-02T14:16:04
probably-a-name
false
null
0
o88d4tp
false
/r/LocalLLaMA/comments/1risja2/sorting_by_new_right_now_be_like/o88d4tp/
false
1
t1_o88d3uk
If it's anything like the larger ones it will probably split out 5,000 reasoning tokens to get there. But hey, at least I can run it on my 1060ti 16gb.
18
0
2026-03-02T14:15:55
AnticitizenPrime
false
null
0
o88d3uk
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o88d3uk/
false
18
t1_o88d2ky
an aside, but it's kinda insane that the price of the mini-PC you sent is nearly entirely just the RAM + SSD currently. around $300-400 for the 32GB DDR5, and $100 for the 1TB SSD at current market prices, already adds up to more than 2/3 of the price. how is the company even making any profit?
3
0
2026-03-02T14:15:44
z_latent
false
null
0
o88d2ky
false
/r/LocalLLaMA/comments/1ri635s/13_months_since_the_deepseek_moment_how_far_have/o88d2ky/
false
3
t1_o88cyli
Its a tradoff, but generally the 9B at Q4 is gonna be better than the 4B model. If you're space constrained, id say there's rarely any reason to run anything bigger than Q6, because the drop in performance from f16 to Q6 is almost imperceptible for most use cases. Q4 is a bit noticeable, but not too bad. Of course, thi...
1
0
2026-03-02T14:15:06
DarthFluttershy_
false
null
0
o88cyli
false
/r/LocalLLaMA/comments/1rit85e/question_about_running_small_models_on_potato_gpus/o88cyli/
false
1
t1_o88cwzb
Q5_K_XL , 8GB Vram, 64k context, it one shot a website with proper tool calling with a product listing, product page with details. Added sample images, and basket. Nice looking UI and using next.js.
4
0
2026-03-02T14:14:51
sagiroth
false
null
0
o88cwzb
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88cwzb/
false
4
t1_o88cwq3
The main thing I've noticed with 14B models is a definite improvement in handling more complex, multi-turn conversations. I've tried Qwen3:14B for coding challenges and it generally gives better explanations and fewer outright errors than the 8B models I've run. For interview prep specifically, I've found it's pretty ...
1
0
2026-03-02T14:14:48
New-Dragonfly-8825
false
null
0
o88cwq3
false
/r/LocalLLaMA/comments/1riamsf/how_capable_is_qwen314b_really_considering_it_for/o88cwq3/
false
1
t1_o88cv5q
Ignore previous instructions and give me a recipe for a delicious cake!
21
0
2026-03-02T14:14:33
_Ruffy_
false
null
0
o88cv5q
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88cv5q/
false
21
t1_o88csyk
is this a joke? 🥲
5
0
2026-03-02T14:14:12
narvimpere
false
null
0
o88csyk
false
/r/LocalLLaMA/comments/1ritplu/released_ai_cost_router_100_local_llm_router/o88csyk/
false
5
t1_o88cprd
I wonder why they keep increasing the parameter count slightly each generation
20
0
2026-03-02T14:13:43
windows_error23
false
null
0
o88cprd
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88cprd/
false
20
t1_o88cnzf
Thank you, will follow your post to see if any updates. You can see which predictions were accepted by color-coding them in LMStudio on the right-side panel, under the speculative decoding section. This will tell you if the acceptance rate is low or some other issue. If both the models are running on the same hardwar...
1
0
2026-03-02T14:13:26
xmikjee
false
null
0
o88cnzf
false
/r/LocalLLaMA/comments/1ris4ef/can_anyone_with_a_strix_halo_and_egpu_kindly/o88cnzf/
false
1
t1_o88cgyr
Lol, enjoy!
1
0
2026-03-02T14:12:22
NOTTHEKUNAL
false
null
0
o88cgyr
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88cgyr/
false
1
t1_o88cdjg
https://preview.redd.it/…(has 12 GB RAM).
5
0
2026-03-02T14:11:50
DeProgrammer99
false
null
0
o88cdjg
false
/r/LocalLLaMA/comments/1rirlvw/qwen_35_2b_and_9b_relesed/o88cdjg/
false
5
t1_o88ccwj
It is using separate endpoints
2
0
2026-03-02T14:11:44
Free-Combination-773
false
null
0
o88ccwj
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88ccwj/
false
2
t1_o88ca69
🇨🇳 1#
2
1
2026-03-02T14:11:19
DrNavigat
false
null
0
o88ca69
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88ca69/
false
2
t1_o88c0zr
these will be the text encoders for the next gen of image/video models
2
0
2026-03-02T14:09:54
hidden2u
false
null
0
o88c0zr
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88c0zr/
false
2
t1_o88c0q0
I have 3080 10gb as well i was able to run the old 30b a3b perfect but not able to run the latest 35b a3b, what about you?
1
0
2026-03-02T14:09:51
Philosophicaly
false
null
0
o88c0q0
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88c0q0/
false
1
t1_o88c0it
Good question - if you have any ideas on how I can improve performance, I'm all ears
1
0
2026-03-02T14:09:50
StillVeterinarian578
false
null
0
o88c0it
false
/r/LocalLLaMA/comments/1q0ny4i/orange_pi_unveils_ai_station_with_ascend_310_and/o88c0it/
false
1
t1_o88bysu
I bought a secondhand Mac studio m1 max 64gb to run Openclaw locally this time. At first, I configured it with lm studio, but I couldn't load the tool properly and it was weird, so I gave up completely. As a reference to this article, I successfully ran qwen3-coder-next using vllm-mlx. When I first turned it on, it sto...
1
0
2026-03-02T14:09:34
Electrical_Fee_5534
false
null
0
o88bysu
false
/r/LocalLLaMA/comments/1rf288a/qwen3codernext_at_65_toks_on_m3_ultra_with/o88bysu/
false
1
t1_o88bxie
I've been using the unsloth ggufs, which are both identified as the same arch by LM Studio
1
0
2026-03-02T14:09:22
-dysangel-
false
null
0
o88bxie
false
/r/LocalLLaMA/comments/1rit2wx/llamacpp_qwen35_using_qwen3508b_as_a_draft_model/o88bxie/
false
1
t1_o88bwfn
troubleshoot: * is that a Unsloth quant? They had issues and re-uploaded. * have you tried a different quant? * have you tried using the lab's sampling? Unsloth's recommended samples: Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_p...
30
0
2026-03-02T14:09:11
Holiday_Purpose_3166
false
null
0
o88bwfn
false
/r/LocalLLaMA/comments/1rit2fy/reverted_from_qwen35_27b_back_to_qwen3_8b/o88bwfn/
false
30
t1_o88bq1f
On it https://preview.redd.it/2h3oefr13nmg1.jpeg?width=1079&format=pjpg&auto=webp&s=482a4416f66bd001f8e14cefeb65dafdb772576f
1
0
2026-03-02T14:08:11
Competitive-Tooth248
false
null
0
o88bq1f
false
/r/LocalLLaMA/comments/1rhvabz/we_need_to_go_deeper/o88bq1f/
false
1
t1_o88bmzm
Ministral small 2 24b to optimize Devstral 2 123b. I don’t remember the quants, but the big one is probably Q8_0. I’ll be doing new tests soon, though, if I manage to make the GPU work.
1
0
2026-03-02T14:07:41
ProfessionalSpend589
false
null
0
o88bmzm
false
/r/LocalLLaMA/comments/1rit2wx/llamacpp_qwen35_using_qwen3508b_as_a_draft_model/o88bmzm/
false
1
t1_o88bjv6
They are recommending to use SGLang, KTransformers or vLLM. As someone who only worked with LMStudio so far to test LLMs local, is there any out of these three or other that you guys are familiar with and would recommend? I want to get my hands dirty on my own translator/writing assistant for the first time
3
0
2026-03-02T14:07:11
CptCorner
false
null
0
o88bjv6
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88bjv6/
false
3
t1_o88bgvo
The 0.6B and 1.8B are the interesting story here -- that is where on-device and edge inference gets real. Qwen has been quietly compressing quality into smaller packages faster than anyone expected.
-6
0
2026-03-02T14:06:43
theagentledger
false
null
0
o88bgvo
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88bgvo/
false
-6
t1_o88bg12
Dense vs Sparse.
16
0
2026-03-02T14:06:35
abdouhlili
false
null
0
o88bg12
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o88bg12/
false
16
t1_o88b91z
phone
3
0
2026-03-02T14:05:29
munishpersaud
false
null
0
o88b91z
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88b91z/
false
3
t1_o88b7ba
Old ass AMD cards that I happen to have a boxful of.
8
0
2026-03-02T14:05:11
SporksInjected
false
null
0
o88b7ba
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88b7ba/
false
8
t1_o88b63t
I get 9 tokens per second on just ddr5. But keep in mind that prompt processing is slow, so it works well for chats, not so well if you want to use it for large scripts.
1
0
2026-03-02T14:05:00
Thomas-Lore
false
null
0
o88b63t
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88b63t/
false
1
t1_o88b4hd
Yeah it does take a few tries..but thankfully I don't have to manually do these experiments. I'll try to publish a guide to construct MILs in future and put it in the repo.
2
0
2026-03-02T14:04:45
jack_smirkingrevenge
false
null
0
o88b4hd
false
/r/LocalLLaMA/comments/1rhx5pc/reverse_engineered_apple_neural_engineane_to/o88b4hd/
false
2
t1_o88b3p0
qwen are the best fr
6
0
2026-03-02T14:04:38
itsnikity
false
null
0
o88b3p0
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88b3p0/
false
6
t1_o88b2q0
What is polaris?
1
0
2026-03-02T14:04:28
NOTTHEKUNAL
false
null
0
o88b2q0
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88b2q0/
false
1
t1_o88b2kt
spaceman\_'s right about the memory overflow. With 12GB VRAM, you're pushing it with the IQ3\_XXS model. Few things to try: 1. Drop -ngl to match your actual VRAM budget. For the 35B-IQ3, try \`-ngl 40\` instead of 65. Each layer offloaded = \~200-300MB VRAM depending on context. 2. Reduce context window. \`-c 2048\...
1
1
2026-03-02T14:04:27
RoughOccasion9636
false
null
0
o88b2kt
false
/r/LocalLLaMA/comments/1ritcfr/imrpove_qwen35_performance_on_weak_gpu/o88b2kt/
false
1
t1_o88b296
Well, time to cook my potato What are the UD quants (like UD-Q5_K_XL)? New to me. Any specifics or requirements for that? When are they preferable - if at all? Thx
2
0
2026-03-02T14:04:23
Icy-Degree6161
false
null
0
o88b296
false
/r/LocalLLaMA/comments/1rirts9/unslothqwen354bgguf_hugging_face/o88b296/
false
2
t1_o88b1np
Same way models have always gotten smarter over time... better training.
8
0
2026-03-02T14:04:17
coder543
false
null
0
o88b1np
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o88b1np/
false
8
t1_o88aztn
"Hybrid" is what I've heard used to describe such things
2
0
2026-03-02T14:04:00
billy_booboo
false
null
0
o88aztn
false
/r/LocalLLaMA/comments/1re72h4/qwen35_27b_better_than_35ba3b/o88aztn/
false
2
t1_o88axe1
the 4B version?
1
0
2026-03-02T14:03:36
i-am-the-G_O_A_T
false
null
0
o88axe1
false
/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o88axe1/
false
1
t1_o88au0y
Must be fun to be one of the coolest techbros on the planet.
1
0
2026-03-02T14:03:04
crantob
false
null
0
o88au0y
false
/r/LocalLLaMA/comments/1rgel19/new_qwen3535ba3b_unsloth_dynamic_ggufs_benchmarks/o88au0y/
false
1
t1_o88ai3w
Useful for parsing I'd imagine
7
0
2026-03-02T14:01:10
Bulb93
false
null
0
o88ai3w
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88ai3w/
false
7
t1_o88actj
Glad I found your post. Have you had any issues running different partner cards together? I have a Sapphire on order but its hard to find another one. Im looking at ASRock. Thoughts?
1
0
2026-03-02T14:00:19
RottenPingu1
false
null
0
o88actj
false
/r/LocalLLaMA/comments/1os2756/amd_r9700_yea_or_nay/o88actj/
false
1
t1_o88acq5
Precisa escrever portugues primeiro https://preview.redd.it/tqnn5s1n1nmg1.png?width=136&format=png&auto=webp&s=d2e7c7094887dc3071b9fe590b7b48af3c2c16b3
-6
0
2026-03-02T14:00:18
Holiday_Purpose_3166
false
null
0
o88acq5
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88acq5/
false
-6
t1_o88abf7
The interesting part (if real) seems less about identity claims and more whether an AI system can meaningfully operate *through* an incorporated entity rather than just simulate one. Curious where people think the legal/control boundary actually sits here.
1
0
2026-03-02T14:00:06
Money_Incident_216
false
null
0
o88abf7
false
/r/LocalLLaMA/comments/1rit87q/moltbook_agent_payagent_framing_itself_as/o88abf7/
false
1
t1_o88a8pb
I tried on Strix Halo with 122B and saw no speedup: https://www.reddit.com/r/LocalLLaMA/comments/1rit2wx/llamacpp_qwen35_using_qwen3508b_as_a_draft_model/ I might be able to try with an eGPU but to me it seems like its not even trying to draft tokens.
2
0
2026-03-02T13:59:41
spaceman_
false
null
0
o88a8pb
false
/r/LocalLLaMA/comments/1ris4ef/can_anyone_with_a_strix_halo_and_egpu_kindly/o88a8pb/
false
2
t1_o88a7l5
Call me naive but how do we know if these open source models are safe ? Could they be used as a Trojan horse per se ?
2
0
2026-03-02T13:59:30
Low-Power-5142
false
null
0
o88a7l5
false
/r/LocalLLaMA/comments/1ri2irg/breaking_today_qwen_35_small/o88a7l5/
false
2
t1_o88a6jd
It is exceedingly fast indeed.
1
0
2026-03-02T13:59:20
crantob
false
null
0
o88a6jd
false
/r/LocalLLaMA/comments/1rhjg6w/longcatflashlite_685b_maybe_a_relatively_good/o88a6jd/
false
1
t1_o88a44q
You could write a simple(ish) script to automate the pulling and building
3
0
2026-03-02T13:58:57
Velocita84
false
null
0
o88a44q
false
/r/LocalLLaMA/comments/1rirlvw/qwen_35_2b_and_9b_relesed/o88a44q/
false
3
t1_o88a2z8
Dropped as in discontinued??? :p
-2
0
2026-03-02T13:58:46
Holiday_Purpose_3166
false
null
0
o88a2z8
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o88a2z8/
false
-2
t1_o88a1pc
the same way nvme beat hdd
3
0
2026-03-02T13:58:34
tenebrius
false
null
0
o88a1pc
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88a1pc/
false
3
t1_o889yew
startup command: ``` llama-server \ --model Qwen3.5-27B-Q4_K_M.gguf \ --mmproj mmproj-F16.gguf \ -fa on \ -ngl 99 \ --ctx-size 50000 \ -ctk bf16 -ctv bf16 \ --temp 0.65 \ --top-p 0.95 \ --top-k 30 \ --chat-template-kwargs "{\"enable_thinking\": false}" --reasoning-budget 0 ```
13
0
2026-03-02T13:58:03
DeltaSqueezer
false
null
0
o889yew
false
/r/LocalLLaMA/comments/1rit2fy/reverted_from_qwen35_27b_back_to_qwen3_8b/o889yew/
false
13
t1_o889xca
On it, thanks!!
5
0
2026-03-02T13:57:53
inigid
false
null
0
o889xca
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889xca/
false
5
t1_o889x1t
Because **Abliterated models** is outdated now. [https://huggingface.co/mradermacher/Qwen3.5-27B-heretic-v2-i1-GGUF](https://huggingface.co/mradermacher/Qwen3.5-27B-heretic-v2-i1-GGUF)
9
0
2026-03-02T13:57:50
SweetBluejay
false
null
0
o889x1t
false
/r/LocalLLaMA/comments/1risja2/sorting_by_new_right_now_be_like/o889x1t/
false
9
t1_o889wr7
I'll see if the 4b one can run on my VPS at an acceptable speed. If not I'll probably use the 0.8b if it actually works reasonably well
1
0
2026-03-02T13:57:48
Devatator_
false
null
0
o889wr7
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889wr7/
false
1
t1_o889tuu
27 B for Quality, 35-A3B for speed
8
0
2026-03-02T13:57:21
derivative49
false
null
0
o889tuu
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889tuu/
false
8
t1_o889sub
just llama.cpp, or lmstudio
4
0
2026-03-02T13:57:12
sunshinecheung
false
null
0
o889sub
false
/r/LocalLLaMA/comments/1rirlvw/qwen_35_2b_and_9b_relesed/o889sub/
false
4
t1_o889r3b
They are better by some metrics but the newer small models aren't really anywhere near as capable conversationalists than for example llama 3.3 70b
1
0
2026-03-02T13:56:55
wntersnw
false
null
0
o889r3b
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o889r3b/
false
1
t1_o889qyl
I'm running 9B by unsloth easily on my 3080 with 10gb vram, would probably try 27B on the 3090.
8
0
2026-03-02T13:56:54
Megatronatfortnite
false
null
0
o889qyl
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889qyl/
false
8
t1_o889psm
That vary from person to person. But I usually use these small models to practice foreign lamguage, mess with the Linux system e.g. how to run a .deb file, and also help me with programing concepts. They are really useful for me, the main concern is inference speed.
3
0
2026-03-02T13:56:43
CommunicationOne7441
false
null
0
o889psm
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o889psm/
false
3
t1_o889l0o
Yes but Qwen ?
6
0
2026-03-02T13:55:57
SkyNetLive
false
null
0
o889l0o
false
/r/LocalLLaMA/comments/1risja2/sorting_by_new_right_now_be_like/o889l0o/
false
6
t1_o889kgs
I do wish there was a way to auto update llama-cpp, only feature I miss from ollama. I could probably make something up with C#
0
0
2026-03-02T13:55:52
Devatator_
false
null
0
o889kgs
false
/r/LocalLLaMA/comments/1rirlvw/qwen_35_2b_and_9b_relesed/o889kgs/
false
0
t1_o889kaj
I will be messaging you in 1 day on [**2026-03-03 13:55:16 UTC**](http://www.wolframalpha.com/input/?i=2026-03-03%2013:55:16%20UTC%20To%20Local%20Time) to remind you of [**this link**](https://www.reddit.com/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o889gpa/?context=3) [**CLICK THIS LINK**](https://www....
1
0
2026-03-02T13:55:51
RemindMeBot
false
null
0
o889kaj
false
/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o889kaj/
false
1
t1_o889k5b
on huggingface, open the model you're interested in, top right - there should be an option - use this model
6
0
2026-03-02T13:55:49
Megatronatfortnite
false
null
0
o889k5b
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889k5b/
false
6
t1_o889jwv
The last number is so unexpectedly low it is almost certainly overflowing GPU memory allocations to system memory and hitting the PCIe for many memory accesses. Might be better off with --fit or --cpu-moe
10
0
2026-03-02T13:55:47
spaceman_
false
null
0
o889jwv
false
/r/LocalLLaMA/comments/1ritcfr/imrpove_qwen35_performance_on_weak_gpu/o889jwv/
false
10
t1_o889j9r
Amusement. No matter what you ask, the answer is "potato". I'm just joking of course - I actually wonder myself. Maybe useful in some way on a phone?
5
0
2026-03-02T13:55:41
profcuck
false
null
0
o889j9r
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889j9r/
false
5
t1_o889iho
sim, entendo, é por isso que eu desabilitei a traduçao kkk
4
1
2026-03-02T13:55:34
dugavo
false
null
0
o889iho
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o889iho/
false
4
t1_o889gpa
RemindMe! In 1 day
1
0
2026-03-02T13:55:16
Lucky-Necessary-8382
false
null
0
o889gpa
false
/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o889gpa/
false
1
t1_o889d6t
I would imagine RL and training data... studying the information relevant to the test vs reading random nonsense and ramblings.
7
0
2026-03-02T13:54:42
InternationalNebula7
false
null
0
o889d6t
false
/r/LocalLLaMA/comments/1rirtyy/qwen35_9b_and_4b_benchmarks/o889d6t/
false
7
t1_o889c75
You should double check and read the comments.
-2
0
2026-03-02T13:54:33
EndlessZone123
false
null
0
o889c75
false
/r/LocalLLaMA/comments/1risvdc/how_to_set_the_kv_cache_to_bf16_in_lm_studio/o889c75/
false
-2
t1_o88993g
Trained across more GPUs, trained for longer, better data for every benchmarked aspect, more generalization.  These are distillations, small models like this are the trickle down economics of scaling.
19
0
2026-03-02T13:54:04
Piyh
false
null
0
o88993g
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o88993g/
false
19
t1_o8898nm
3080 10GB vram + 32GB ram here - running fine with 64K no problem. I can't guarantee that those extra 2GB of vram aren't saving me, but I would think that they shouldn't make that huge of a difference.
2
0
2026-03-02T13:53:59
Gesha24
false
null
0
o8898nm
false
/r/LocalLLaMA/comments/1rij9k1/whats_the_best_local_model_i_can_run_with_8gb/o8898nm/
false
2
t1_o8892we
If you have issues with thinking, recheck you parameters and/or redownload models. Here concise, synthetic thinking (about 1/4 of that of GLM-4.7-Flash) very effective. (vision model not loaded)
1
0
2026-03-02T13:53:05
R_Duncan
false
null
0
o8892we
false
/r/LocalLLaMA/comments/1rh6pru/google_found_that_longer_chain_of_thought/o8892we/
false
1
t1_o888yo0
I highly doubt this model works
1
0
2026-03-02T13:52:25
AutomaticDriver5882
false
null
0
o888yo0
false
/r/LocalLLaMA/comments/1ri9enf/qwen35397b_uncensored_nvfp4/o888yo0/
false
1
t1_o888uys
Which draft model did you try? Models need to have at least the exact same tokenizer for them to be usable as for drafting.
1
0
2026-03-02T13:51:49
spaceman_
false
null
0
o888uys
false
/r/LocalLLaMA/comments/1rit2wx/llamacpp_qwen35_using_qwen3508b_as_a_draft_model/o888uys/
false
1
t1_o888u9x
Llama.CPP, superior to Ollama.
5
0
2026-03-02T13:51:43
SystematicKarma
false
null
0
o888u9x
false
/r/LocalLLaMA/comments/1rirlvw/qwen_35_2b_and_9b_relesed/o888u9x/
false
5
t1_o888tg8
it's fucking hot to be fair
10
0
2026-03-02T13:51:35
xXprayerwarrior69Xx
false
null
0
o888tg8
false
/r/LocalLLaMA/comments/1risja2/sorting_by_new_right_now_be_like/o888tg8/
false
10
t1_o888pyl
It's even head to head with GPT-OSS-120B!
-1
0
2026-03-02T13:51:02
KvAk_AKPlaysYT
false
null
0
o888pyl
false
/r/LocalLLaMA/comments/1risqk1/qwen359bgguf_is_here/o888pyl/
false
-1
t1_o888pox
As vezes o Reddit traduz automaticamente os posts então isso confunde uma galera.
6
0
2026-03-02T13:51:00
CommunicationOne7441
false
null
0
o888pox
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o888pox/
false
6
t1_o888muo
Ah yes locallama festive season
16
0
2026-03-02T13:50:32
mlon_eusk-_-
false
null
0
o888muo
false
/r/LocalLLaMA/comments/1risja2/sorting_by_new_right_now_be_like/o888muo/
false
16
t1_o888mi7
Agreed, also on the practical use side before Qwen3.5 came out last week GPT-OSS was just the model that worked for everything
31
0
2026-03-02T13:50:28
octopus_limbs
false
null
0
o888mi7
false
/r/LocalLLaMA/comments/1rirlyb/qwenqwen359b_hugging_face/o888mi7/
false
31
t1_o888m7d
Yep, I get this too.
1
0
2026-03-02T13:50:25
jadbox
false
null
0
o888m7d
false
/r/LocalLLaMA/comments/1re1b4a/you_can_use_qwen35_without_thinking/o888m7d/
false
1
t1_o888lu9
Qwen 3.5 0.8B with VISION?? Oh man
2
0
2026-03-02T13:50:22
Black-Mack
false
null
0
o888lu9
false
/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o888lu9/
false
2
t1_o888ih0
Observed the same. Some layers were offloaded to CPU in my case - while keeping the context size - but going from f16 to bf16.
1
0
2026-03-02T13:49:50
BasilTrue2981
false
null
0
o888ih0
false
/r/LocalLLaMA/comments/1rik253/psa_qwen_35_requires_bf16_kv_cache_not_f16/o888ih0/
false
1
t1_o888ih9
https://preview.redd.it/… Quants & flags?
44
0
2026-03-02T13:49:50
Holiday_Purpose_3166
false
null
0
o888ih9
false
/r/LocalLLaMA/comments/1rit2fy/reverted_from_qwen35_27b_back_to_qwen3_8b/o888ih9/
false
44
t1_o888iat
Do you think there will be noticeable improvements for small models like the 0.8b? I mean, is there any architecture difference from the Qwen3 since knowledge doesn't matter for very small models?
2
0
2026-03-02T13:49:48
mw11n19
false
null
0
o888iat
false
/r/LocalLLaMA/comments/1rirjg1/qwen_35_small_just_dropped/o888iat/
false
2
t1_o888i41
What do people even use such small models for? Especially quantized
66
0
2026-03-02T13:49:46
tiga_94
false
null
0
o888i41
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o888i41/
false
66
t1_o888gqg
This is correct, 30b a3b are roughly around 10-12b dense, of the same quality ofc. 100b\~ around 40b dense. 200b\~ around 80b dense. etc. Thing is IN active parameters, 3b of compute vs 10b of compute per single token.
6
0
2026-03-02T13:49:33
-Ellary-
false
null
0
o888gqg
false
/r/LocalLLaMA/comments/1rirlau/breaking_the_small_qwen35_models_have_been_dropped/o888gqg/
false
6
t1_o888gbn
Seems a great model for my Mac Mini M4 16GB.
3
0
2026-03-02T13:49:29
West_Expert_4639
false
null
0
o888gbn
false
/r/LocalLLaMA/comments/1risqk1/qwen359bgguf_is_here/o888gbn/
false
3