[whatmodel.sh](../../README.md) / Qwen3.5 4B
# Qwen3.5 4B
Alibaba · released 2026-03 · 4.66B params · apache-2.0
**score 60.0** ±0.5 · #201 of 732 · from 130 benchmark results across 17 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Qwen3.5 4B among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆█●▅▅▄▄▂▁ #170 of 313 11% 13
coding ▁▁▁▃▅▇●▇▅▄▃▁ #242 of 469 13% 9
chat & writing ▁▁▁▃▆▄▆█●▇▇▃ #166 of 402 0.9% 1
math ▁▁▃▅█▆●█▄▂▁▁ #154 of 431 22% 13
vision & documents ▁▁▂▂▁▃▄▂●▇█▄ #90 of 153 10% 21
web search ▅·▅▅█·●▃▃·▅▃ #6 of 15* 1.2% 2
agent tasks · #170 of 313 · 11% of its score
Beats [o1 (2024-12-17)](o1-2024-12-17.md), [o3-mini](o3-mini.md), [DeepSeek R1 0528](deepseek-r1-0528.md) here, though they rank higher overall. Trails [SmolLM3 3B](smollm3-3b.md), [Jamba 1.5 Mini](jamba-1-5-mini.md), [Jamba 1.5 Large](jamba-1-5-large.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
BFCL V4 (December 2025 archive) 56.8% ▄·█·▄▄▄●·▄·▄ beats 6/8 medium · avg 43% 3.1% at its level · several runs
SWE-bench Multilingual 27.7% ▁··▁●▁▅██▂▂▁ beats 4/65 hard · avg 18% 1.0% at its level
SWE-bench Verified 33.6% ▄··▄▄█··▄·●█ beats 6/8 very hard · avg 11% 1.0% sharp · at its level · noisy
SWE-bench Pro (split unspecified) 28.2% █▂···▂···▄·● beats 8/8 very hard · avg 3.9% 0.8% sharp · noisy
BrowseComp-ZH 39.6% ▄▄·▄▄▄·█···● beats 7/8 very hard · avg 13% 0.8% at its level · noisy
OSWorld-Verified (maker) 35.6% ●▃·▂▅▂▃▂▅▃▆▆ beats 0/27 hard · avg 19% 0.8% at its level · small run
MCPMark 12.5% ▃·····●··▃█▃ beats 1/8 very hard · avg 5.5% 0.7% typical
τ³-bench Banking 5.5% ▂▂·▅●▂▃▃▃▆▂▅ beats 6/24 very hard · avg 4.0% 0.5% blunt · too hard for it
BFCL v3 64.1% ▁▁▁▁▁▃▅█▆▅●▁ beats 165/171 medium · avg 55% 0.5% blunt · at its level · several runs
τ²-bench aggregate (maker) 79.9% ▁▁▁·▁▁▃▃▃●█▁ beats 24/39 medium · avg 40% 0.4% at its level · noisy
τ³-bench 6.7% ▃·●·█·▃··▃·▃ beats 1/7 very hard · avg 13% 0.4% typical
BrowseComp 14.3% ▅···█·●···▃▃ beats 5/7 very hard · avg 11% 0.4% noisy
τ²-bench Telecom 87.7% ▄█▄········● beats 4/5 hard · avg 25% 0.4% sharp · at its level · noisy
coding · #242 of 469 · 13% of its score
Trails [QwQ-32B-Preview](qwq-32b-preview.md), [InternLM2.5 20B Chat](internlm2-5-20b-chat.md), [Ministral 3 3B Reasoning 2512](ministral-3-3b-reasoning.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LiveCodeBench v6 (window unspecif… 56.4% ▃····█·▃●▃·▃ beats 6/8 hard · avg 33% 4.8% at its level · several runs
LiveCodeBench v6 (Feb–May 2025) (… 55.8% ▁▁▃▁●▅▄▇▅▅▃█ beats 10/38 medium · avg 41% 1.6% at its level · several runs
LiveCodeBench Pro 2025Q2 (Medium) 7.0% █···▂··▄●··▂ beats 7/8 very hard · avg 2.0% 1.5% sharp · too hard for it · noisy
LiveCodeBench Pro 2025Q2 (Easy) 58.3% ▄···█▄··▄█●▄ beats 7/8 hard · avg 33% 1.2% at its level
SWE-bench Multilingual 27.7% ▁··▁●▁▅██▂▂▁ beats 4/65 hard · avg 18% 1.0% at its level
SWE-bench Verified 33.6% ▄··▄▄█··▄·●█ beats 6/8 very hard · avg 11% 1.0% sharp · at its level · noisy
SWE-bench Pro (split unspecified) 28.2% █▂···▂···▄·● beats 8/8 very hard · avg 3.9% 0.8% sharp · noisy
SciCode (level unspecified) 24.0% █·····██●███ beats 3/6 hard · avg 17% 0.7% at its level
MultiPL-E 51.0% ▃··▄·●▅▇▄█▃▄ beats 8/31 medium · avg 54% 0.3% at its level
chat & writing · #166 of 402 · 0.9% of its score
Beats [MiniMax-M2](minimax-m2.md), [MiniCPM5 2B](minicpm5-2b.md), [Granite 4.2 30B](granite-4-2-30b.md) here, though they rank higher overall. Trails [LFM2.5 1.2B Thinking](lfm2-5-1-2b-thinking.md), [LFM2.5 8B-A1B](lfm2-5-8b-a1b.md), [LFM2.5 2.6B](lfm2-5-2-6b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
IFEval (aggregation unspecified) 90.2% ▄▄····▄··█●█ beats 5/8 easy · avg 80% 0.9% small run · several runs
math · #154 of 431 · 22% of its score
Beats [MiMo-V2.5 (310B-A15B)](mimo-v2-5.md), [MiniMax-M2.7](minimax-m2-7.md), [MiniMax-M2.1](minimax-m2-1.md) here, though they rank higher overall. Trails [Nemotron 3 Nano 30B-A3B](nemotron-3-nano-30b-a3b.md), [Qwen3-VL 8B Thinking](qwen3-vl-8b-thinking.md), [Olmo 3 32B Think](olmo-3-32b-think.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
HMMT February 2026 64.0% ▅··▃▃·▃··▃·● beats 8/8 hard · avg 32% 6.1% sharp · at its level · several runs
AIME 2025 78.8% █▄▄▄▄····●·▄ beats 6/8 medium · avg 42% 3.4% sharp · at its level · several runs
AIME 2026 82.7% ▄·▄█·▄▄···●▄ beats 6/8 medium · avg 53% 3.1% sharp · several runs
OTIS mock AIME 2024–2025 55.8% ▃▃▄▃▃▃●▇█▃▂▅ beats 84/197 very hard · avg 7.2% 1.8% sharp · at its level · noisy
IMO-AnswerBench 68.5% ▄▄···▄·▄··●▄ beats 4/6 hard · avg 32% 1.7% at its level
MATH-Vision (maker) 74.6% ▁▁·▁··▄▁▃●▇█ beats 12/24 medium · avg 54% 1.4% at its level · small run · several runs
HMMT February 2025 (maker) 74.0% ▁▁▁▃▁▅●▃██▁▄ beats 12/31 medium · avg 49% 1.0% at its level
HMMT November 2025 (maker) 76.8% ▁▁····▁●▃█▇▁ beats 4/23 medium · avg 62% 0.9% typical
PolyMATH (maker) 51.1% ▁···▃▁·▇●▄▄▄ beats 13/23 hard · avg 34% 0.7% at its level · small run
MATH-500 99.0% ██▄▄▄··▄···● beats 8/8 very easy · avg 89% 0.6% maxed out
MathVista (testmini) (maker) 85.1% ▁··▁▄▄▆▇▄▅●▄ beats 39/48 medium · avg 68% 0.6% small run
DynaMath (maker) 83.3% ▂·····▂··●▄▆ beats 4/10 easy · avg 69% 0.5% small run
We-Math (maker) 75.4% ●··········▃ beats 2/3 medium · avg 58% 0.4% at its level · small run
vision & documents · #90 of 153 · 10% of its score
Beats [DeepSeek V4 Flash Vision (Exp)](deepseek-v4-flash-vision-exp.md), [GPT-5 mini](gpt-5-mini.md), [GLM-4.6V](glm-4-6v.md) here, though they rank higher overall.
test result vs the others in its run difficulty counts why
MATH-Vision (maker) 74.6% ▁▁·▁··▄▁▃●▇█ beats 12/24 medium · avg 54% 1.4% at its level · small run · several runs
ChartArena chart parsing (English… 33.5% ▁·▁·▃▃▁●▃█·▄ beats 8/22 hard · avg 27% 1.2% blunt · at its level
BEAR-Bench (bilingual RU/EN enter… 48.6% ▃▃·▅·▃·●·█▅▅ beats 5/14 hard · avg 24% 0.7% at its level
CharXiv (reasoning questions) (ma… 70.8% ▂▅▃▅▅▆▆●▆█▆▂ beats 58/101 medium · avg 46% 0.7% at its level · small run
MathVista (testmini) (maker) 85.1% ▁··▁▄▄▆▇▄▅●▄ beats 39/48 medium · avg 68% 0.6% small run
MMMU (val) (maker) 77.6% ▁····▁▁▃█▆●▅ beats 43/53 medium · avg 60% 0.6% at its level · small run
MMMU Pro (maker) 66.3% ▂▃▁▃▆▇▃●▇██▅ beats 28/52 medium · avg 48% 0.6% at its level · small run
OmniDocBench 1.5 overall (higher … 86.2% ▂··▂▂▂··●▂█▅ beats 4/13 easy · avg 73% 0.5% small run
MMStar (maker) 78.3% ▃▅▂▃▂▂▆██●█▅ beats 25/35 medium · avg 65% 0.5% at its level · small run
CountBench (maker) 96.3% ▂▂▃▂·▂▂▃▃·●█ beats 12/18 very easy · avg 92% 0.4% maxed out · small run
PMC-VQA (maker) 55.5% ▃···▃··●▃·▃█ beats 2/7 medium · avg 42% 0.4% at its level · small run
SimpleVQA (maker) 43.4% ▃·▃·●·▃·█▅▃▃ beats 2/10 hard · avg 34% 0.3% blunt · at its level · small run
MMBench-EN dev v1.1 (maker) 89.4% █▄···▂··●▂▄▆ beats 7/14 easy · avg 82% 0.3% blunt · small run
ScreenSpot-Pro (maker) 60.3% ▂▃▅▆●█▅▂···▂ beats 11/24 medium · avg 48% 0.3% blunt · at its level · small run
HallusionBench (maker) 65.0% ▂▁·▂▂▂▃▅▃●▅▅ beats 19/33 medium · avg 56% 0.2% blunt · at its level · small run
RealWorldQA (maker) 79.5% ▁▃▁▇▃▇▇▅●▁▅▅ beats 29/39 easy · avg 72% 0.2% blunt · small run
SpaceCast-Bench predictive spatia… 33.7% ▃·●▃▅▃··▅··█ beats 1/11 very hard · avg 33% 0.2% noisy
CC-OCR (maker) 76.7% ▁·▂··▂▂●▃▂▆█ beats 9/26 easy · avg 72% 0.2% blunt · small run
OCRBench (maker) 85.0% ▁·▃▁▆▅●█▅█▁▃ beats 15/37 easy · avg 83% 0.1% blunt · small run
AI2D test (maker) 89.6% ▁▁▅▇▃▇·●·█·▃ beats 19/28 easy · avg 85% 0.1% blunt · small run
PureDocBench document parsing (Av… 69.8% ▂·▁·▂▂▅●·▁▂▂ beats 14/22 medium · avg 62% 0.0% blunt · at its level
web search · #6 of 15 · 1.2% of its score
Beats [Qwen3.5 9B](qwen3-5-9b.md), [Gemma 4 12B](gemma-4-12b.md) here, though they rank higher overall. Trails [Spark-X2.5 1.7B](spark-x2-5-1-7b.md), [Spark-X2.6 1.7B](spark-x2-6-1-7b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
BrowseComp-ZH 39.6% ▄▄·▄▄▄·█···● beats 7/8 very hard · avg 13% 0.8% at its level · noisy
BrowseComp 14.3% ▅···█·●···▃▃ beats 5/7 very hard · avg 11% 0.4% noisy
other tests (46)
test result vs the others in its run difficulty counts why
GPQA Diamond 76.8% ▁▁▁▁▂▄▅▅█●█▄ beats 189/322 hard · avg 46% 9.0% sharp · at its level · several runs
MMLU-Redux 88.7% ▂···█··▂▂··● beats 8/8 easy · avg 78% 3.3% several runs
MMLU-Pro 78.0% ▂··▂█·▂▂···● beats 8/8 medium · avg 61% 3.3% several runs
IFBench (aggregation unspecified) 59.0% ▅····▃▃·●·▃▃ beats 5/8 medium · avg 39% 2.3% at its level · several runs
Humanity's Last Exam 8.6% ▄▄·▄··▄·●·▄█ beats 4/7 very hard · avg 3.8% 2.3% typical
LongBench v2 47.3% ▃······▅▅▃·● beats 8/8 very hard · avg 35% 2.2% at its level · several runs
GPQA (subset unspecified) 67.2% █···▃···●·▃▃ beats 4/7 hard · avg 44% 1.9% sharp · at its level
OJBench 24.8% ▃····▃▅··█●▃ beats 7/8 very hard · avg 12% 1.7% several runs
KINA (Knowledge Index of Noah's A… 28.5% ▁▁▂▁▃▄●▆▄▅▂▃ beats 16/39 very hard · avg 21% 1.3% typical
C-Eval (maker) 85.1% ▁···▁▁▂▁▂▂●█ beats 14/33 medium · avg 73% 1.3% typical
SuperGPQA 52.8% ▃·▃·▃▅█····● beats 8/8 hard · avg 32% 1.3% at its level · several runs
MMMLU (maker) 76.1% ▂▁▁▂▃▂▄▄●██▃ beats 23/46 medium · avg 63% 1.1% at its level
MedXPertQA MM (maker) 42.9% ▂···▂·▃··●▂█ beats 4/13 very hard · avg 30% 1.0% sharp · at its level · small run
KOPA-Bench (pass@1) 17.6% ▄·●··▄··█▄·█ beats 1/8 very hard · avg 14% 1.0% at its level
INCLUDE (maker) 71.0% ▂····▂▂▄▄●█▄ beats 17/36 medium · avg 56% 1.0% at its level · small run
EuroEval Life in the UK (English … 86.1% ▂·▂▂·▂·●▃·█▃ beats 4/14 easy · avg 80% 0.9% typical
Global PIQA (maker) 78.9% ▃▃··▃··●▃▅█▅ beats 3/11 medium · avg 71% 0.8% at its level
MultiChallenge (maker) 49.0% ▁·▁·▁▃▆●▃▂█▅ beats 13/30 hard · avg 31% 0.8% at its level
MAXIFE (maker) 78.0% ▂·▂·▂···●·▄█ beats 4/11 medium · avg 64% 0.8% typical
EuroEval HellaSwag (English), val… 76.9% ▂··▂▂·▂▆●█·▄ beats 9/15 medium · avg 63% 0.7% at its level
MMLU-ProX (maker) 71.5% ▁·▁▁▁▂▁●▃█▅▄ beats 8/36 medium · avg 56% 0.7% at its level
VitaBench (maker) 22.0% ▄··▂▄●▆█▄▄▄▂ beats 7/21 very hard · avg 11% 0.7% typical
GSM-Plus 85.2% ▁··▁▁▁·▃▁▄●▁ beats 15/16 easy · avg 84% 0.6% blunt · several runs
BFCL multi-turn (version not stat… 49.1% ▄▄···▄·●█▄▄▄ beats 3/8 hard · avg 26% 0.5% at its level
NOVA-63 (maker) 54.3% ▃▃·▃···▃●▃▃█ beats 6/11 medium · avg 39% 0.5% blunt · at its level
Multi-IF (aggregation unspecified) 73.6% ▄▄···▄·▄·▄●█ beats 6/8 medium · avg 60% 0.5% at its level · several runs
VideoMMMU (maker) 74.1% ▂···▂···●▂█▆ beats 2/10 medium · avg 62% 0.5% at its level · small run
VLMs Are Blind (maker) 92.6% ▂·▂·····●··█ beats 2/8 very easy · avg 87% 0.4% maxed out · small run
Claw-Eval average score 62.3% ▄·▄··●▄▄▄▄▄▄ beats 2/9 medium · avg 53% 0.4% blunt · at its level
LVBench (maker) 66.4% ▇·█▅●▄▄··▃▁▁ beats 17/28 medium · avg 54% 0.4% at its level · small run
Video-MME (with subtitles) (maker) 83.5% ▁▁▁▁▁·▁▃▁●▁█ beats 9/16 medium · avg 73% 0.4% small run
SLAKE (maker) 76.1% ▄·····▄●·██▄ beats 2/7 medium · avg 64% 0.4% small run
MMVU (maker) 64.9% ▃···▃···●▃▅█ beats 2/8 medium · avg 50% 0.4% at its level · small run
LingoQA (maker) 74.4% ●······▄·█·█ beats 0/5 medium · avg 61% 0.4% small run
MLVU (maker) 82.8% ▂···▂···●▄▄█ beats 2/10 easy · avg 72% 0.4% small run
ERQA (maker) 54.0% ▃··▃··●·▃█▅▅ beats 3/12 medium · avg 38% 0.4% at its level · small run
RefSpatial-Bench (maker) 54.6% ▃·▃····●▃▅▅█ beats 2/10 medium · avg 40% 0.4% at its level · small run
MMLongBench-Doc (maker) 54.2% ▃···▃···●▅█▅ beats 2/9 medium · avg 39% 0.4% at its level · small run
AndroidWorld (maker) 58.6% ▁·▁▁▂▃●▅▂▂·▂ beats 14/24 medium · avg 41% 0.4% at its level · small run
ZEROBench subquestions (maker) 26.3% ▄·▄···●·▄██▄ beats 2/8 hard · avg 17% 0.4% at its level · small run
EmbSpatial-Bench (maker) 81.3% ▁·····▁·●·▃█ beats 2/10 easy · avg 70% 0.4% small run
Video-MME (without subtitles) (ma… 76.9% ▁▃▁▃▁·█▄●▅·▇ beats 18/27 medium · avg 67% 0.3% blunt · at its level · small run
DeepPlanning (maker) 17.6% ▂·····▂··●█▄ beats 2/9 very hard · avg 7.1% 0.2% noisy
MVBench (maker) 71.2% ▁··▃▄▁▃▄●▄█▄ beats 12/26 medium · avg 66% 0.2% blunt · at its level · small run
LIBRA Mini (Russian long context,… 39.1% ▄···▄●·██··▄ beats 3/8 medium · avg 38% 0.1% blunt · at its level
WMT24++ (maker) 66.6% ▁·▁▁·▁·●▂▆█▃ beats 4/27 easy · avg 66% 0.1% blunt · at its level
source results share of the score
Qwen/Qwen3.5-9B model card (submitted /… 62 34% ▇▇▇▇
openbmb/MiniCPM5-2B model card 19 20% ▇▇
XHToken/Spark-X2.5-4B model card 11 14% ▇▇
JetBrains/Mellum2-12B-A2.5B-Instruct mo… 8 8% ▇
Qwen/Qwen3.5-4B model card (submitted /… 4 6% ▇
LiquidAI/LFM2.5-2.6B model card 7 5% ▇
JetBrains/Mellum2-12B-A2.5B-Thinking mo… 7 4% ▇
Epoch AI 1 2% ▇
EuroEval leaderboards 2 2% ▇
Multi-Step Tool-Calling over Korean Ope… 2 2% ▇
Knowledge Index of Noah's Ark 1 1% ▇
ChartArena: Benchmarking Chart Parsing … 1 1% ▇
BEAR-Bench: A Bilingual Enterprise and … 1 1% ▇
LiquidAI/LFM2.5-8B-A1B model card 1 0% ▇
SpaceCast-Bench: Evaluating Predictive … 1 0% ▇
ai-forever LIBRA leaderboard 1 0% ▇
How Far Is Document Parsing from Solved… 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 60.0
Licences: Qwen/Qwen3.5-9B model card ([apache-2.0](../../licenses/Qwen--Qwen3.5-9B.txt)) · openbmb/MiniCPM5-2B model card ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)) · XHToken/Spark-X2.5-4B model card ([Apache-2.0](https://huggingface.co/XHToken/Spark-X2.5-4B/blob/0bcb35678590218655dff3765b9e61c83b35e9c4/README.md)) · JetBrains/Mellum2-12B-A2.5B-Instruct model card ([Apache-2.0](https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Instruct/blob/273d9c1c6882ac8e1562e5b7e043857d35a787b6/README.md)) · Qwen/Qwen3.5-4B model card ([apache-2.0](../../licenses/Qwen--Qwen3.5-4B.txt)) · LiquidAI/LFM2.5-2.6B model card ([LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-2.6B/blob/654f9463ce32b05d0429d76fe1f580b27d4c1ac0/LICENSE)) · JetBrains/Mellum2-12B-A2.5B-Thinking model card ([Apache-2.0](https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Thinking/blob/a7311550557e93cc706ab5dd3d879c1a11703ab4/README.md)) · Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE)) · Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Knowledge Index of Noah's Ark ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · ChartArena: Benchmarking Chart Parsing across Languages ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LiquidAI/LFM2.5-8B-A1B model card ([LFM Open License v1.0](../../licenses/cards-gap-NOTICE.txt)) · SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · ai-forever LIBRA leaderboard ([MIT](../../licenses/hf-leaderboard-spaces-MIT.txt)) · How Far Is Document Parsing from Solved? PureDocBench: A Source-Traceable Benchmark across Clean ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/))
## How to run it
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Qwen3.5 4B (Q4_K_M, needs ~2.7 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF/tree/f9f88ac3e234be915e23811a6d28ea287bdb927e))
```sh
llama cli -hf lmstudio-community/Qwen3.5-4B-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf lmstudio-community/Qwen3.5-4B-GGUF:Q4_K_M
```
Other versions: Q8_0 (~4.5 GB) `llama cli -hf lmstudio-community/Qwen3.5-4B-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Qwen3.5 4B (Q4_K_M, needs ~2.7 GB, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF/tree/f9f88ac3e234be915e23811a6d28ea287bdb927e))
```sh
lms get https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~4.5 GB) `lms get https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF@Q8_0 -y` · 4bit (~3 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3.5-4B-MLX-4bit -y` · 8bit (~5.1 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3.5-4B-MLX-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Qwen3.5 4B (Q4_K_M, needs ~3.3 GB, from [ollama.com](https://ollama.com/library/qwen3.5/tags))
```sh
ollama run qwen3.5:4b
```
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Qwen3.5 4B (4bit, needs ~3 GB, Mac only, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3.5-4B-MLX-4bit/tree/c43ee1d65576a5d98de1e8405cac93c371a655c1))
```sh
mlx_lm.chat --model lmstudio-community/Qwen3.5-4B-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model lmstudio-community/Qwen3.5-4B-MLX-4bit
```
Other versions: 8bit (~5.1 GB, Mac only) `mlx_lm.chat --model lmstudio-community/Qwen3.5-4B-MLX-8bit`
Model repository and documentation: [huggingface.co/Qwen](https://huggingface.co/Qwen/Qwen3.5-4B)
## Quantization evidence
How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent. They don't describe every download of that precision.
precision test full → quantized kept
AutoRound AutoScheme mixed-pr… MATH-500 (EvalScope via llama-ser… 0.79 → 0.754 95.4%
AutoRound AutoScheme mixed-pr… GPQA Diamond (EvalScope via llama… 0.47 → 0.384 81.7%
AutoRound AutoScheme mixed-pr… MATH-500 (EvalScope via llama-ser… 0.79 → 0.72 91.1%
AutoRound AutoScheme mixed-pr… GPQA Diamond (EvalScope via llama… 0.47 → 0.424 90.3%
AutoRound AutoScheme mixed-pr… MATH-500 (EvalScope via llama-ser… 0.79 → 0.786 99.5%
AutoRound AutoScheme mixed-pr… GPQA Diamond (EvalScope via llama… 0.47 → 0.419 89.2%
AutoRound AutoScheme mixed-pr… MATH-500 (EvalScope via llama-ser… 0.79 → 0.804 101.8%
AutoRound AutoScheme mixed-pr… GPQA Diamond (EvalScope via llama… 0.47 → 0.419 89.2%
FP8 W8A8 dynamic (E4M3; weigh… IFEval prompt-level strict 0-shot… 0.88 → 0.884 100.5%
FP8 W8A8 dynamic (E4M3; weigh… GSM8K-Platinum 0-shot (lm-evaluat… 0.942 → 0.945 100.3%
FP8 W8A8 dynamic (E4M3; weigh… MATH-500 0-shot (lighteval, vLLM;… 0.846 → 0.847 100.1%
FP8 W8A8 dynamic (E4M3; weigh… AIME 2025 0-shot (lighteval, vLLM… 0.85 → 0.85 100.0%
FP8 W8A8 dynamic (E4M3; weigh… MMLU-Pro 0-shot (lm-evaluation-ha… 0.793 → 0.791 99.7%
FP8 W8A8 dynamic (E4M3; weigh… GPQA Diamond 0-shot (lighteval, v… 0.768 → 0.763 99.3%
GGUF Q3_K_M (AutoRound-genera… MATH-500 (EvalScope via llama-ser… 0.79 → 0.772 97.7%
GGUF Q3_K_M (AutoRound-genera… GPQA Diamond (EvalScope via llama… 0.47 → 0.404 86.1%
GGUF Q3_K_M (unsloth file as … MATH-500 (EvalScope via llama-ser… 0.79 → 0.776 98.2%
GGUF Q3_K_M (unsloth file as … GPQA Diamond (EvalScope via llama… 0.47 → 0.404 86.0%
GGUF Q4_K_M (AutoRound-genera… MATH-500 (EvalScope via llama-ser… 0.79 → 0.792 100.3%
GGUF Q4_K_M (AutoRound-genera… GPQA Diamond (EvalScope via llama… 0.47 → 0.455 96.8%
GGUF Q4_K_M (repeat 5) (AutoR… MATH-500 (EvalScope via llama-ser… 0.79 → 0.796 100.7%
GGUF Q4_K_M (repeat 5) (AutoR… GPQA Diamond (EvalScope via llama… 0.47 → 0.432 92.0%
GGUF Q4_K_M (repeat 5) (unslo… MATH-500 (EvalScope via llama-ser… 0.79 → 0.791 100.1%
GGUF Q4_K_M (repeat 5) (unslo… GPQA Diamond (EvalScope via llama… 0.47 → 0.442 94.2%
GGUF Q4_K_M (unsloth file as … MATH-500 (EvalScope via llama-ser… 0.79 → 0.794 100.5%
GGUF Q4_K_M (unsloth file as … GPQA Diamond (EvalScope via llama… 0.47 → 0.384 81.7%
GGUF UD-IQ2_M (unsloth file a… MATH-500 (EvalScope via llama-ser… 0.79 → 0.594 75.2%
GGUF UD-IQ2_M (unsloth file a… GPQA Diamond (EvalScope via llama… 0.47 → 0.227 48.4%
INT4 W4A16 (GPTQ, symmetric, … GSM8K-Platinum 0-shot (lm-evaluat… 0.945 → 0.945 100.0%
INT4 W4A16 (GPTQ, symmetric, … MATH-500 0-shot (lighteval, vLLM;… 0.845 → 0.837 99.1%
INT4 W4A16 (GPTQ, symmetric, … MMLU-Pro 0-shot (lm-evaluation-ha… 0.793 → 0.775 97.7%
INT4 W4A16 (GPTQ, symmetric, … IFEval prompt-level strict 0-shot… 0.883 → 0.861 97.5%
INT4 W4A16 (GPTQ, symmetric, … GPQA Diamond 0-shot (lighteval, v… 0.796 → 0.731 91.8%
INT4 W4A16 (GPTQ, symmetric, … AIME 2025 0-shot (lighteval, vLLM… 0.822 → 0.733 89.2%
INT8 W8A8 (weights static per… GSM8K-Platinum 0-shot (lm-evaluat… 0.945 → 0.942 99.7%
INT8 W8A8 (weights static per… MMLU-Pro 0-shot (lm-evaluation-ha… 0.793 → 0.79 99.6%
INT8 W8A8 (weights static per… IFEval prompt-level strict 0-shot… 0.883 → 0.877 99.3%
INT8 W8A8 (weights static per… MATH-500 0-shot (lighteval, vLLM;… 0.845 → 0.839 99.3%
INT8 W8A8 (weights static per… AIME 2025 0-shot (lighteval, vLLM… 0.822 → 0.812 98.8%
INT8 W8A8 (weights static per… GPQA Diamond 0-shot (lighteval, v… 0.796 → 0.785 98.6%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/qwen3-5-4b) · [JSON](https://whatmodel.sh/api?format=json&all=1)