whatmodel.sh / Qwen3 235B-A22B (Apr 2025, hybrid)

Qwen3 235B-A22B (Apr 2025, hybrid)

+ contribute

Qwen3 235B-A22B (Apr 2025, hybrid) is an open-weights language model from Alibaba, released April 2025. Its published size is 235B parameters (22B active per token) and a 131k-token context window. Licence: apache-2.0.

59.8±0.5#197 of 719 LLMsfrom 104 benchmark results across 20 sources

$ how it compares

The LLMs ranked either side of Qwen3 235B-A22B (Apr 2025, hybrid), out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
235	ERNIE 4.5 300B-A47B Baidu	60±1	local	~97GB	2025-06
236	Mistral Medium 3 (2505) Mistral	60±1	comparison only		2025-05
237	Qwen3 235B-A22B (Apr 2025, hybrid) Alibaba	60±1	api $1.23/M +local	~76GB	2025-04
238	Claude 3.5 Sonnet (2024-10-22) Anthropic	60±1	comparison only		2024-10
239	Granite 4.2 8B IBM	60±1	api $0.11/M +local	~6.6GB	2026-08
	more from Alibaba
18	Qwen3.8 Max (2.4T-A95B) Alibaba	78±1	api $3.00/M +local	~380GB	2026-08
34	Qwen3.8 Max (0902) Alibaba	76±1	api $3.00/M		2026-09
37	Qwen3.8-Flash-Next Alibaba	75±1	local	~59GB	2026-08

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #187 of 300 · 25% of its score · 6 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 0528 here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Mellum2 12B-A2.5B Instruct here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
τ²-bench Telecom +5	22.1% ✎	██··█·██··██ beats 1/6	very hard · avg 8.1%	▇▇▇▇ 10%	sharp · at its level · several runs
τ-bench Retail lab +3	65.2% ✎	▁····▃·▂▇█▂▆ beats 14/28	medium · avg 40%	▇▇ 4.1%	at its level · small run · several runs
τ²-bench Retail +4	51.4% ✎	▁····▂▁▃▆█▃▁ beats 18/44	medium · avg 38%	▇▇ 3.9%	at its level · several runs
τ²-bench Airline +4	36.8% ✎	▁▁·▃▁▃▂▁█▅▂▃ beats 3/42	hard · avg 29%	▇ 3.0%	at its level · several runs
τ-bench Airline lab +3	32.0% ✎	▁▁▁·▅▄▁█▇▁▁▇ beats 6/28	hard · avg 22%	▇ 2.5%	at its level · small run · several runs
BFCL v3 +3	67.9% ✎	▁▁▁▁▁▃▅█▅▄▁▁ beats 104/114	medium · avg 52%	▇ 1.2%	blunt · at its level · several runs

coding▁▁▁▃▆██▇▆▄▃▁ #149 of 458 · 14% of its score · 5 tests

Beats Gemini 2.5 Pro, Gemini 3.1 Flash-Lite, Qwen3-Coder-Next (80B-A3B) here, though they rank higher overall. Trails Ministral 3 8B Reasoning 2512, Phi-4-reasoning 14B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Aider polyglot (225 exercises, two attempts) +1	59.6% ✎	▂·▂▃▅▂▃█▆▄▂▂ beats 34/49	very hard · avg 8.5%	▇▇▇ 7.8%	sharp · at its level · several runs
LiveCodeBench v6 (Feb–May 2025) lab +3	55.7% ✎	▁▁▃▁█▅▄▇▅▅▃█ beats 9/38	hard · avg 26%	▇ 2.9%	at its level · small run · several runs
LiveCodeBench official (Aug 1, 2024–May 1, 2025)	65.9% ✎	▃▃▅▃··▃▅█▅▅▃ beats 9/15	hard · avg 25%	▇ 1.5%	at its level
WebApp1K-Duo-React pass@1 (onekq-ai)	32.9% ✎	▁··▁·▃▁▁▆█▇█ beats 3/45	hard · avg 32%	▇ 1.0%	at its level
MultiPL-E lab	79.3% ✎	▃··▄·▅▅▇▄█▃▄ beats 25/31	easy · avg 66%	▇ 0.6%	small run

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #134 of 393 · 11% of its score · 5 tests

Beats o3-mini, EXAONE 4.5 33B, INTELLECT-3 here, though they rank higher overall. Trails LFM2.5 8B-A1B, LFM2.5 2.6B, Spark-X2.5 1.7B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena-Hard v2 aggregate lab +2	61.5% ✎	▁▁··▄▃▁▃█▁▁▇ beats 10/22	hard · avg 27%	▇▇ 3.9%	sharp · at its level · small run · several runs
Arena-Hard v2.0 hard prompts (style control, Gemini-2.5 judge)	58.4% ✎	▂·▃▅▅·▃█▆▂▂▂ beats 17/22	very hard · avg 12%	▇ 2.9%	sharp · at its level
LMArena (text)	1375.2 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 217/379	avg 1268	▇ 1.9%	blunt
IFEval (aggregation unspecified) lab +2	83.4% ✎	▁▁▃▁▂▄▄▆█▃▂▁ beats 66/141	easy · avg 77%	▇ 1.4%	several runs
Arena-Hard v2.0 creative writing (GPT-4.1 + Gemini-2.5 judges)	73.5% ✎	▆▂▂▆▂▂▃█▂▃·▂ beats 20/22	hard · avg 19%	▇ 0.5%	sharp · at its level · noisy

math▁▂▃▅▆▅▄█▄▁▁▁ #202 of 398 · 15% of its score · 5 tests

Trails Olmo 3 7B Instruct, Phi-4-mini-reasoning 3.8B, DeepSeek R1 Distill Qwen 7B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2025 lab +3	81.5% ✎	▁·▁▁▁▂▅▆█▇▅▁ beats 88/133	medium · avg 55%	▇▇ 4.8%	sharp · small run · several runs
HMMT 2025 (edition unspecified) lab +2	62.5% ✎	▃█▇▅▇▅▁·▁··▁ beats 15/32	hard · avg 33%	▇▇ 4.7%	sharp · at its level · small run · several runs
MATH level 5	68.9% ✎	▂▂▅▆▄█▆▃▅▂▃▄ beats 59/92	hard · avg 32%	▇ 3.4%	sharp
PolyMATH lab +2	54.7% ✎	▁···▃▁·▇█▄▄▄ beats 15/23	hard · avg 34%	▇ 1.8%	at its level · small run · several runs
THINK-Bench final-answer accuracy (reasoning models)	94.9% ✎	▄█·▄····▄▄·█ beats 7/7	easy · avg 76%	▇ 0.3%	typical

other tests23 not in a task · 36% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA (subset unspecified) lab +2	71.1% ✎	▂▁▂▂▄▃▆▃▅█▄▃ beats 78/104	hard · avg 50%	▇▇ 4.7%	sharp · at its level · small run · several runs
MMLU-Pro lab +3	82.8% ✎	▁▁·▁▁▁▃▂▄▅█▁ beats 153/188	medium · avg 59%	▇▇ 3.8%	at its level · several runs
GPQA Diamond	70.7% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 106/208	hard · avg 40%	▇ 3.5%	sharp · at its level
INCLUDE lab +2	78.7% ✎	▂····▂▂▄▄▄█▄ beats 23/36	medium · avg 59%	▇ 3.0%	at its level · several runs
MMLU-Redux lab +2	92.7% ✎	▁▁▃▁▂▃▄▅▇▅█▄ beats 48/63	easy · avg 81%	▇ 2.8%	several runs
Creative Writing v3 lab +2	84.6% ✎	▂····▂▆··▆█▆ beats 7/14	easy · avg 70%	▇ 1.9%	small run · several runs
MMLU-ProX lab +2	80.0% ✎	▁·▁▁▁▂▁▃▂█▅▄ beats 23/35	medium · avg 59%	▇ 1.9%	at its level · several runs
SuperGPQA lab +2	60.7% ✎	▂▁▂▂▅▆▃▇▇▇█▃ beats 35/46	medium · avg 35%	▇ 1.7%	at its level · small run · several runs

15 more tests

test	result	vs the others in its run	difficulty	counts	why
WritingBench lab +2	80.3% ✎	▁·▁▂▃··▅█▂▂▁ beats 4/20	easy · avg 66%	▇ 1.5%	several runs
Multi-IF (aggregation unspecified) lab +2	71.9% ✎	▁▁▁▁▁▁▃▄█▄▁▂ beats 14/41	medium · avg 60%	▇ 1.5%	blunt · at its level · several runs
OJBench lab +1	25.6% ✎	▁···▁▃·▅▄▅██ beats 14/26	very hard · avg 13%	▇ 1.4%	small run · several runs
LiveBench 2024-11-25 lab +2	77.1% ✎	▄··▂·▂·▂█▆▂▂ beats 11/13	medium · avg 62%	▇ 1.4%	small run · several runs
EQ-Bench 3 (emotional intelligence, judge Elo)	1215.9 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 32/62	avg 1044	▇ 1.2%	typical
BRIDGE clinical text, zero-shot average (87 tasks)	41.6% ✎	▁▁▁▁▁▂▂▃▄▆█▃ beats 84/97	hard · avg 32%	▇ 1.1%	blunt · at its level
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	29.4% ✎	▁▁▂▁▃▄█▆▄▅▂▃ beats 18/39	very hard · avg 22%	▇ 1.1%	blunt
CPTUB average (SpeakLeash, 0-5 judge scale)	78.2% ✎	▁▂▂▂▃█▃▆▅▄▂▃ beats 76/95	easy · avg 66%	▇ 1.0%	typical
Humanity's Last Exam lab	11.8% ✎	▁▁▁▃▆▆▇██▆▁▃ beats 24/104	very hard · avg 4.8%	▇ 0.7%	small run
ChineseSimpleQA lab	60.8% ✎	▄··▄▄▄··▄··█ beats 1/6	medium · avg 52%	▇ 0.5%	at its level · small run
ZebraLogic lab +1	37.7% ✎	█·▁▁▁▇·▄▄▁▄▁ beats 5/24	medium · avg 47%	▇ 0.4%	at its level · noisy · small run · several runs
SimpleQA lab	12.2% ✎	▄▄▄▆█▆███·▂█ beats 15/32	very hard · avg 10%	▇ 0.3%	noisy · small run
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy	85.9% ✎	▃·▁▁▂▁▂▅█▇▄▁ beats 85/100	medium · avg 68%	▇ 0.2%	noisy
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy	95.6% ✎	▁▁▁▁▁▁▁▂▄█▄▃ beats 143/146	easy · avg 83%	▇ 0.1%	noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	90.7% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 53/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $1.22/M · $0.7 in / $2.8 out per 1M tokens · 131k context · www.alibabacloud.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Qwen3 235B-A22B (Apr 2025, hybrid) · needs ~71 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/Qwen3-235B-A22B-GGUF:UD-Q2_K_XL
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Qwen3-235B-A22B-GGUF:UD-Q2_K_XL

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Qwen3 235B-A22B (Apr 2025, hybrid) · needs ~124.9 GB, Mac only · from huggingface.co/Qwen (official)
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/Qwen/Qwen3-235B-A22B-MLX-4bit -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Qwen3 235B-A22B (Apr 2025, hybrid) · needs ~141.5 GB · from ollama.com
chat in the terminal
ollama run qwen3:235b-a22b-q4_K_M

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Qwen3 235B-A22B (Apr 2025, hybrid) · needs ~124.9 GB, Mac only · from huggingface.co/Qwen (official)
chat in the terminal
mlx_lm.chat --model Qwen/Qwen3-235B-A22B-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model Qwen/Qwen3-235B-A22B-MLX-4bit

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Qwen/Qwen3-235B-A22B-Thinking… 22%, Qwen/Qwen3-30B-A3B-Thinking-2… 20%, Qwen/Qwen3-30B-A3B-Instruct-2… 20% and 17 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Qwen/Qwen3-235B-A22B-Thinking-2507 model card · lab-reported	24	▇▇▇▇▇▇▇ 22%	apache-2.0
Qwen/Qwen3-30B-A3B-Thinking-2507 model card · lab-reported	23	▇▇▇▇▇▇ 20%	apache-2.0
Qwen/Qwen3-30B-A3B-Instruct-2507 model card · lab-reported	25	▇▇▇▇▇▇ 20%	apache-2.0
Qwen/Qwen3-235B-A22B-Instruct-2507 model card · lab-reported	11	▇▇ 8%	apache-2.0
Epoch AI — internal evaluations	2	▇▇ 7%	CC BY 4.0
Aider polyglot benchmark runs	1	▇▇ 5%	Apache-2.0
Arena-Hard-Auto v2.0-Preview leaderboard (repository README)	2	▇ 3%	Apache-2.0
HyperCLOVA X 32B Think (NAVER Cloud, arXiv:2601.03286v1)	3	▇ 3%	CC-BY-4.0
LMArena leaderboard dataset	1	▇ 2%	CC BY 4.0
Moonshot AI Kimi-K2-Instruct model card (instruction-model evaluation table)	1	▇ 2%	Modified MIT License (Kimi K2)
LiveCodeBench official website results (2025-08-01)	1	▇ 2%	CC-BY-SA-4.0
EQ-Bench 3 canonical leaderboard results	1	▇ 1%	MIT
BRIDGE Medical Leaderboard (Harvard Medical School / Brigham and Women's YLab)	1	▇ 1%	MIT
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 1%	CC BY 4.0
WebApp1K models leaderboard (onekq-ai)	1	▇ 1%	MIT
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)	1	▇ 1%	MIT
Berkeley Function Calling Leaderboard result archive	1	▇ 0%	Apache-2.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇ 0%	MIT
THINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models (arXiv:2505.22113v1)	1	▇ 0%	CC-BY-4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 59.8

$ quantization evidence · 18 measurements

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.

precision	test	full → quantized	kept
FP8 W8A8 dynamic (E4M3; weights static per-channel, activations dynamic per-token; llm-compressor FP8_DYNAMIC, compressed-tensors); lm_head and MoE router gates unquantized	GSM8K 5-shot strict-match (OpenLLM v1 configuration; lm-evaluation-harness, vLLM)	0.742 → 0.75	101.0%	source ↗
FP8 W8A8 dynamic (E4M3; weights static per-channel, activations dynamic per-token; llm-compressor FP8_DYNAMIC, compressed-tensors); lm_head and MoE router gates unquantized	OpenLLM v1 average (6 tasks: MMLU 5-shot, ARC-Challenge 25-shot, GSM8K 5-shot strict-match, HellaSwag 10-shot, Winogrande 5-shot, TruthfulQA mc2 0-shot; lm-evaluation-harness, vLLM)	0.738 → 0.738	100.0%	source ↗
FP8 W8A8 dynamic (E4M3; weights static per-channel, activations dynamic per-token; llm-compressor FP8_DYNAMIC, compressed-tensors); lm_head and MoE router gates unquantized	MMLU 5-shot (OpenLLM v1 configuration; lm-evaluation-harness, vLLM)	0.848 → 0.846	99.8%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	GPQA (subset not stated in card), thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.711 → 0.719	101.1%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	GPQA (subset not stated in card), non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.629 → 0.628	99.8%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	MMLU-Redux, non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.892 → 0.89	99.8%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	MMLU-Redux, thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.927 → 0.92	99.2%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	LiveBench 2024-11-25, non-thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.625 → 0.611	97.8%	source ↗
GPTQ INT4 (w4, group_size 128, sym, desc_act=false; checkpoint Qwen/Qwen3-235B-A22B-GPTQ-Int4)	LiveBench 2024-11-25, thinking mode (Qwen-reported in quantized model card; harness, sampling and number of runs not stated)	0.771 → 0.751	97.4%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	GSM8K (OpenLLM v1 configuration, 5-shot; lm-evaluation-harness, vLLM)	0.851 → 0.864	101.5%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	IFEval (OpenLLM v2 leaderboard configuration; lm-evaluation-harness, vLLM)	0.882 → 0.882	100.0%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	OpenLLM v1 average (6 tasks: arc_challenge, gsm8k, hellaswag, mmlu, truthfulqa_mc2, winogrande; lm-evaluation-harness (neuralmagic fork), vLLM)	0.788 → 0.787	99.9%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	OpenLLM v2 average as published (5 tasks listed: BBH, MMLU-Pro, MuSR, IFEval, GPQA; Math-Hard not included; lm-evaluation-harness, vLLM)	0.565 → 0.564	99.8%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	MMLU-Pro (OpenLLM v2 leaderboard configuration, 5-shot; lm-evaluation-harness, vLLM)	0.582 → 0.58	99.6%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	MMLU (OpenLLM v1 configuration, 5-shot; lm-evaluation-harness, vLLM)	0.864 → 0.858	99.3%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	MATH-500 0-shot (lighteval, vLLM)	0.95 → 0.942	99.2%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	Reasoning average (2 tasks: GPQA Diamond 0-shot, MATH-500 0-shot; lighteval, vLLM)	0.836 → 0.817	97.7%	source ↗
NVFP4 W4A4 (FP4 E2M1 weights and activations, group 16; llm-compressor QuantizationModifier NVFP4, compressed-tensors); lm_head, self-attention projections and MoE router gates left unquantized	GPQA Diamond 0-shot (lighteval, vLLM)	0.722 → 0.692	95.8%	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json