[whatmodel.sh](../../README.md) / Qwen3.6 35B-A3B
# Qwen3.6 35B-A3B
Alibaba · released 2026-04 · 35.5B params (3B active) · apache-2.0
**score 66.8** ±0.5 · #114 of 732 · from 89 benchmark results across 11 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Qwen3.6 35B-A3B among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆█▄▅●▄▄▂▁ #95 of 313 22% 12
coding ▁▁▁▃▅▇█▇●▄▃▁ #96 of 469 19% 7
math ▁▁▃▅█▆▄●▄▂▁▁ #132 of 431 37% 13
vision & documents ▁▁▂▂▁▃▄▂▆●█▄ #58 of 153 5.8% 15
web search ▁▁●▅▄█▃▄·▁·▁ #76 of 84* 1.7% 2
agent tasks · #95 of 313 · 22% of its score
Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [GPT-5.1](gpt-5-1.md), [o3-pro](o3-pro.md) here, though they rank higher overall. Trails [Hy3 preview (295B-A21B)](hy3-preview.md), [MiniMax-M2.7](minimax-m2-7.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified 64.4% ▄▄·····▄▄▄●█ beats 5/7 hard · avg 16% 6.8% sharp · at its level · several runs
SWE-bench Multilingual 57.0% ▄·▄·····█▄●▄ beats 5/7 hard · avg 18% 3.6% at its level · several runs
Terminal-Bench 2.1 34.8% ▄······█▄▄●▄ beats 5/7 very hard · avg 7.2% 2.6% sharp · at its level
SWE-bench Pro (split unspecified) 40.6% ▄···▄···▄▄●█ beats 5/7 very hard · avg 6.2% 2.2% sharp · at its level
SWE-bench Pro public set, Qwen-re… 49.5% ▄······●█·▄▄ beats 1/5 hard · avg 19% 1.6% at its level · several runs
K-BrowseComp-Verified (Korean web… 12.0% ▃····●··▃█·▃ beats 3/8 very hard · avg 3.6% 1.1% typical
Terminal-Bench 2.0 (maker) 51.5% ▂▂▅▅██●█▆▂▃▃ beats 21/36 hard · avg 15% 1.0% at its level · small run
MCP-Atlas (public set) (maker) 62.8% ▂·▁▂▂·▁▃●▆▆▂ beats 15/39 hard · avg 30% 0.9% at its level · small run
Toolathlon (Tool Decathlon) (make… 26.9% ▂▂●▂▄▄▆▂█▆██ beats 3/27 very hard · avg 11% 0.8% at its level · small run
MCPMark (maker) 37.0% ▃·····█··▃●▃ beats 7/8 very hard · avg 11% 0.7% at its level · small run
WideSearch (maker) 60.1% ●·▂▆··▂···▂▂ beats 2/10 medium · avg 35% 0.6% at its level · small run
τ³-bench (maker) 67.2% ▁▁▂▁·▁▁·▁▁▂● beats 15/25 medium · avg 40% 0.5% at its level · small run
coding · #96 of 469 · 19% of its score
Beats [Grok 4.20 (0309 reasoning)](grok-4-20.md), [GPT-5.1](gpt-5-1.md), [Grok 4.3](grok-4-3.md) here, though they rank higher overall. Trails [Llama 3.1 Nemotron Ultra 253B v1](llama-3-1-nemotron-ultra-253b.md), [Qwen3 235B-A22B Thinking 2507](qwen3-235b-a22b-thinking-2507.md), [o4-mini](o4-mini.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Verified 64.4% ▄▄·····▄▄▄●█ beats 5/7 hard · avg 16% 6.8% sharp · at its level · several runs
SWE-bench Multilingual 57.0% ▄·▄·····█▄●▄ beats 5/7 hard · avg 18% 3.6% at its level · several runs
SWE-bench Pro (split unspecified) 40.6% ▄···▄···▄▄●█ beats 5/7 very hard · avg 6.2% 2.2% sharp · at its level
Luau Leetcode fully solved (Torpe… 50.0% ▁▁▁▁·▁●▃▅▃█▇ beats 8/49 hard · avg 17% 2.1% sharp · at its level
LiveCodeBench v6 (Feb–May 2025) (… 80.4% ▁▁▃▁█▅▄▇▅▅●█ beats 31/38 medium · avg 48% 2.1% several runs
SWE-bench Pro public set, Qwen-re… 49.5% ▄······●█·▄▄ beats 1/5 hard · avg 19% 1.6% at its level · several runs
SciCode (level unspecified) 37.5% ▄···▄▄█▄·●·▄ beats 6/7 very hard · avg 13% 1.1% at its level
math · #132 of 431 · 37% of its score
Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [MiniMax-M3](minimax-m3.md), [MiMo-V2.5 (310B-A15B)](mimo-v2-5.md) here, though they rank higher overall. Trails [Olmo 3.1 32B Think](olmo-3-1-32b-think.md), [Qwen3-VL 30B-A3B Thinking](qwen3-vl-30b-a3b-thinking.md), [Granite 4.2 8B](granite-4-2-8b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
FrontierMath tiers 1–3 v2 17.5% ▂··▁▁●▄▄█▆▄▃ beats 10/86 very hard · avg 1.2% 9.1% sharp · at its level
MathArena ArXivMath June 2026 (48… 27.1% ▂···●·██▆▆▄▂ beats 1/19 very hard · avg 5.1% 5.1% sharp · at its level
HMMT February 2026 (maker) 83.6% ▃▃▂▅·███●▅▃▅ beats 25/35 medium · avg 36% 4.6% sharp · several runs
MathArena ArXivMath April 2026 (4… 33.8% ▁·····●██▄·▁ beats 2/20 very hard · avg 7.6% 4.4% sharp · at its level
AIME 2026 (maker) 92.7% ▁▁▁▂▂▂▆●▆▂·▁ beats 35/52 medium · avg 59% 4.0% sharp · maxed out · small run · several…
IMO-AnswerBench (maker) 78.9% ▁▁·▁▁▁▄▆●▄▂▃ beats 21/44 medium · avg 36% 3.1% small run · several runs
MathArena ArXivMath May 2026 (40 … 37.5% ▁··●▄█▄▁·▃·▁ beats 1/17 very hard · avg 11% 1.6% at its level
OTIS mock AIME 2024–2025 86.7% ▃▃▄▃▃▃▇▇●▃▂▅ beats 134/197 very hard · avg 7.2% 1.2% sharp · noisy
HMMT February 2025 (maker) 90.7% ▁▁▁▃▁▅▅▃●█▁▄ beats 19/31 easy · avg 65% 0.9% small run
MathArena BrokenArXiv June 2026 (… 3.4% ●··█▂▂▂▆▄··▂ beats 1/14 very hard · avg 1.3% 0.7% sharp · noisy
MathVista (testmini) (maker) 86.4% ▁··▁▄▄▆▇▄▅●▄ beats 44/48 medium · avg 63% 0.7% small run
HMMT November 2025 (maker) 89.1% ▁▁····▁▅▃●▇▁ beats 10/23 easy · avg 68% 0.6% small run
DynaMath (maker) 82.8% ▂·····▂··●▄▆ beats 2/10 medium · avg 63% 0.5% small run
vision & documents · #58 of 153 · 5.8% of its score
Beats [GPT-6 Luna](gpt-6-luna.md), [DeepSeek V4 Flash Vision (Exp)](deepseek-v4-flash-vision-exp.md), [GPT-5](gpt-5.md) here, though they rank higher overall. Trails [Qwen3.6 Flash](qwen3-6-flash.md), [Qwen3.5 35B-A3B](qwen3-5-35b-a3b.md), [Gemma 4 31B](gemma-4-31b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
CharXiv (reasoning questions) (ma… 78.0% ▂▅▃▅▅▆▆▅●█▆▂ beats 72/101 medium · avg 41% 0.8% small run
MathVista (testmini) (maker) 86.4% ▁··▁▄▄▆▇▄▅●▄ beats 44/48 medium · avg 63% 0.7% small run
MMMU (val) (maker) 81.7% ▁····▁▁▃█▆▆● beats 49/53 medium · avg 57% 0.7% small run
OmniDocBench 1.5 overall (higher … 89.9% ▂··▂▂▂··▂▂●▅ beats 10/13 easy · avg 72% 0.6% small run
MMMU Pro (maker) 75.3% ▂▃▁▃▆▇▃▇▇●█▅ beats 40/52 medium · avg 46% 0.6% at its level · small run
MMStar (maker) 80.7% ▃▅▂▃▂▂▆██▆●▅ beats 30/35 medium · avg 62% 0.5% small run
CountBench (maker) 96.1% ▂▂▃▂·▂▂▃▃·●█ beats 11/18 very easy · avg 91% 0.4% maxed out · small run
SimpleVQA (maker) 58.9% ▃·▃·▃·▃·█●▃▃ beats 7/10 hard · avg 34% 0.4% blunt · at its level · small run
MMBench-EN dev v1.1 (maker) 92.8% █▄···▂··▄▂▄● beats 12/14 easy · avg 82% 0.3% blunt · small run
HallusionBench (maker) 69.8% ▂▁·▂▂▂▃▅▃█▅● beats 31/33 medium · avg 54% 0.3% blunt · at its level · small run
RealWorldQA (maker) 85.3% ▁▃▁▇▃▇▇▅█▁▅● beats 37/39 easy · avg 72% 0.2% blunt · small run
CC-OCR (maker) 81.9% ▁·▂··▂▂▃▃▂▆● beats 24/26 easy · avg 72% 0.2% blunt · small run
OCRBench (maker) 90.0% ▁·▃▁▆▅▂█▅●▁▃ beats 32/37 easy · avg 83% 0.2% blunt · small run
AI2D test (maker) 92.7% ▁▁▅▇▃▇·▄·●·▃ beats 22/28 easy · avg 86% 0.1% blunt · small run
PureDocBench document parsing (Av… 67.1% ▂·▁·▂▂●█·▁▂▂ beats 8/22 medium · avg 62% 0.0% blunt · at its level
web search · #76 of 84 · 1.7% of its score
Trails [GPT-4.1 nano](gpt-4-1-nano-2025-04-14.md), [GPT-4o mini (2024-07-18)](gpt-4o-mini-2024-07-18.md), [Mistral Small 3.1 24B (2503)](mistral-small-3-1.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
K-BrowseComp-Verified (Korean web… 12.0% ▃····●··▃█·▃ beats 3/8 very hard · avg 3.6% 1.1% typical
WideSearch (maker) 60.1% ●·▂▆··▂···▂▂ beats 2/10 medium · avg 35% 0.6% at its level · small run
other tests (30)
test result vs the others in its run difficulty counts why
GPQA Diamond 83.8% ▁▁▁▁▂▄▆▄▅●█▅ beats 142/211 hard · avg 44% 6.7% sharp · several runs
Humanity's Last Exam (maker) 21.4% ▁▁▁▄▆▆●██▆▁▃ beats 54/111 very hard · avg 5.1% 2.7% several runs
MMLU-Pro (maker) 85.2% ▁▁·▁▁▁▃▂▄▅●▁ beats 173/188 medium · avg 56% 2.6% several runs
NL2Repo (maker) 29.4% ▄·····●·▄▄·▄ beats 2/5 very hard · avg 8.9% 1.8% at its level · several runs
QwenClawBench (Qwen internal) (ma… 52.6% ▄·▄····▄···● beats 4/4 hard · avg 23% 1.5% at its level · several runs
Mystery Game Puzzles (Epoch) 22.0% █··▂▃▇●█▂▄▂▁ beats 39/77 very hard · avg 9.4% 1.1% sharp · too hard for it · noisy
Claw-Eval Pass^3 (maker) 50.0% █●▄·▄·▄▄·▄·▄ beats 2/8 hard · avg 25% 1.1% at its level · several runs
Chess Puzzles (Epoch) 26.0% █▁▁▁▁▂▃●▃▂▁▁ beats 97/137 very hard · avg 5.8% 1.0% sharp · noisy
SkillsBench 78-task subset (OpenC… 28.7% ▄·····▄···▄● beats 4/4 very hard · avg 11% 0.9% at its level · several runs
RobloxQA v2.0 multiple choice (To… 89.8% ▁▁▃▁●▄▃▃▆▂▂▂ beats 13/49 easy · avg 77% 0.9% typical
VitaBench (maker) 35.6% ▄··▂▄▆▆█▄●▄▂ beats 18/21 very hard · avg 12% 0.8% at its level · small run
C-Eval (maker) 90.0% ▁···▁▁▂▁▂▂▃● beats 22/33 medium · avg 70% 0.8% small run
Claw-Eval average score (maker) 68.7% ▄·▄··█▄▄▄●▄▄ beats 7/9 medium · avg 48% 0.8% blunt · at its level · several runs
MMLU-Redux (maker) 93.3% ▁▁▃▁▂▃▄▅▇▅●▄ beats 54/63 easy · avg 78% 0.7% small run
LVBench (maker) 71.4% ▇·█▅▅●▄··▃▁▁ beats 20/28 medium · avg 52% 0.6% at its level · small run
Next Steps data preparation (bala… 81.7% ▄·▄···█·▄●·▄ beats 5/6 medium · avg 52% 0.5% typical
DeepPlanning (maker) 25.9% ▂·····▂··▄●▄ beats 7/9 very hard · avg 9.5% 0.5% small run
SuperGPQA (maker) 64.7% ▂▁▂▂▅▆▃▇▇▇●▃ beats 40/46 medium · avg 36% 0.5% at its level · small run
ZEROBench subquestions (maker) 34.4% ▄·▄···▄·▄●█▄ beats 5/8 hard · avg 16% 0.5% at its level · small run
DiagnosticIQ industrial maintenan… 69.9% ▁··▃▁▁▁▅█▅●▅ beats 20/24 hard · avg 51% 0.5% blunt · at its level
VideoMMMU (maker) 83.7% ▂···▂···▂▂█● beats 8/10 medium · avg 62% 0.5% small run
VLMs Are Blind (maker) 96.6% ▂·▂·····▃··● beats 4/8 very easy · avg 89% 0.5% maxed out · small run
Video-MME (with subtitles) (maker) 86.6% ▁▁▁▁▁·▁▃▁▁▁● beats 11/16 medium · avg 71% 0.4% small run
ERQA (maker) 61.8% ▃··▃··█·▃●▅▅ beats 6/12 medium · avg 37% 0.4% at its level · small run
Video-MME (without subtitles) (ma… 82.5% ▁▃▁▃▁·█▄▄▅·● beats 23/27 medium · avg 67% 0.4% blunt · small run
MLVU (maker) 86.2% ▂···▂···▂▄▄● beats 7/10 easy · avg 70% 0.4% small run
RefSpatial-Bench (maker) 64.3% ▃·▃····▃▃●▅█ beats 5/10 medium · avg 41% 0.4% at its level · small run
EmbSpatial-Bench (maker) 84.3% ▁·····▁·▁·▃● beats 6/10 easy · avg 68% 0.4% small run
NTCIR SciClaimEval Subtask 1 pair… 66.2% ▃▃···▃·●▃█·▃ beats 3/8 medium · avg 45% 0.4% blunt · at its level
MVBench (maker) 74.6% ▁··▃▄▁▃▄▄▄●▄ beats 19/26 medium · avg 65% 0.2% blunt · at its level · small run
source results share of the score
Qwen/Qwen3.6-27B model card (submitted … 41 35% ▇▇▇▇
Qwen/Qwen3.6-35B-A3B model card (submit… 27 22% ▇▇▇
Epoch AI 5 16% ▇▇
MathArena ArXivMath / BrokenArXiv / ArX… 4 12% ▇
Kwaipilot/KAT-Coder-V2.5-Dev model card 5 10% ▇
Roblox LLM Leaderboard 2 3% ▇
K-BrowseComp: A Web Browsing Agent Benc… 1 1% ▇
Agentic coding without the cloud: evalu… 1 1% ▇
DiagnosticIQ: A Benchmark for LLM-Based… 1 0% ▇
SciTrue: Reliable Scientific Claim Vali… 1 0% ▇
How Far Is Document Parsing from Solved… 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 66.8
Licences: Qwen/Qwen3.6-27B model card ([apache-2.0](../../licenses/Qwen--Qwen3.6-27B.txt)) · Qwen/Qwen3.6-35B-A3B model card ([apache-2.0](../../licenses/Qwen--Qwen3.6-35B-A3B.txt)) · Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets ([CC BY-SA 4.0](../../licenses/matharena-sharealike-NOTICE.txt)) · Kwaipilot/KAT-Coder-V2.5-Dev model card ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)) · Roblox LLM Leaderboard ([MIT](../../licenses/roblox-llm-leaderboard-MIT.txt)) · K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Agentic coding without the cloud: evaluating open-weight LLMs on longitudinal data preparation tasks ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · SciTrue: Reliable Scientific Claim Validation with Frontier and Open Language Models at the NTCIR SciClaimEval Task ([CC BY 4.0](../../licenses/paper-2609.00654v1-CC-BY-4.0.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
### Through an API
api (Deep Infra) · $0.31/M blended · $0.1 in / $0.95 out per 1M tokens · [deepinfra.com](https://deepinfra.com/models)
[price source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/providers/deepinfra/models/Qwen/Qwen3.6-35B-A3B.toml): Standard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ. · [context source](https://github.com/anomalyco/models.dev/blob/a668908cfe7073faa0363699486fc41a3ff52412/models/alibaba/qwen3.6-35b-a3b.toml)
### 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 Qwen3.6 35B-A3B (Q4_K_M, needs ~20.4 GB, from [huggingface.co/ggml-org](https://huggingface.co/ggml-org/Qwen3.6-35B-A3B-GGUF/tree/baec3ebee244827cda0f4557eafa8b28f7545fa6))
```sh
llama cli -hf ggml-org/Qwen3.6-35B-A3B-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf ggml-org/Qwen3.6-35B-A3B-GGUF:Q4_K_M
```
Other versions: Q8_0 (~36.9 GB) `llama cli -hf ggml-org/Qwen3.6-35B-A3B-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 Qwen3.6 35B-A3B (Q4_K_M, needs ~20.4 GB, from [huggingface.co/ggml-org](https://huggingface.co/ggml-org/Qwen3.6-35B-A3B-GGUF/tree/baec3ebee244827cda0f4557eafa8b28f7545fa6))
```sh
lms get https://huggingface.co/ggml-org/Qwen3.6-35B-A3B-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 (~36.9 GB) `lms get https://huggingface.co/ggml-org/Qwen3.6-35B-A3B-GGUF@Q8_0 -y` · 4bit (~20.4 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit -y` · 8bit (~37.7 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-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 Qwen3.6 35B-A3B (Q4_K_M, needs ~21.8 GB, from [ollama.com](https://ollama.com/library/qwen3.6/tags))
```sh
ollama run qwen3.6:35b-a3b
```
### 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 Qwen3.6 35B-A3B (4bit, needs ~20.4 GB, Mac only, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit/tree/0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b))
```sh
mlx_lm.chat --model lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit
```
Other versions: 8bit (~37.7 GB, Mac only) `mlx_lm.chat --model lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit`
Model repository and documentation: [huggingface.co/Qwen](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
## Published speed measurements
Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine.
hardware precision speed engine and settings
AMD Ryzen AI Max+ 395 (Stri… Q4_0 81.3 tok/s llama.cpp 2496f9c14 (build_number…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 76.94 tok/s llama.cpp 2496f9c14 (build_number…
AMD Ryzen AI Max+ 395 (Stri… UD-Q4_K_M 62.56 tok/s llama.cpp 2496f9c14 (build_number…
AMD Ryzen AI Max+ 395 (Stri… UD-Q4_K_M 63.06 tok/s llama.cpp d05fe1d7d (build_number…
AMD Ryzen AI Max+ 395 (Stri… UD-Q4_K_M 52.69 tok/s llama.cpp b1c70e2e5 (build_number…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 72.55 tok/s Ollama 0.31.1 controlled local bi…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 73.19 tok/s Ollama 0.31.2 controlled local bi…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 73.2 tok/s Ollama 0.32.0 controlled local bi…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 50.51 tok/s Ollama 0.23.1, Ollama Vulkan RADV…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 49.05 tok/s Ollama 0.24.0, Ollama Vulkan RADV…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 49.09 tok/s Ollama 0.23.1, Ollama Vulkan RADV…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 50.51 tok/s Ollama 0.21.2, Ollama Vulkan RADV…
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 63.76 tok/s llama.cpp b8460, Vulkan (RADV) · …
AMD Ryzen AI Max+ 395 (Stri… Q4_K_M 60.57 tok/s Ollama 0.31.2 system service, Oll…
Apple M5 Pro (48GB; CPU/GPU… Q4 GGUF (4-bit;… 63.1 tok/s llama.cpp b9960, Metal backend, d…
Apple M5 Pro (48GB; CPU/GPU… MLX 4-bit (grou… 101.9 tok/s mlx-lm 0.31.3 (MLX 0.32.0), defau…
Each speed links to its source, which has the full settings.
## 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
FP8 block-wise (E4M3, 128x128… GPQA Diamond 0-shot (lighteval vi… 0.845 → 0.859 101.6%
FP8 block-wise (E4M3, 128x128… GSM8K-Platinum 0-shot (lm-evaluat… 0.957 → 0.96 100.3%
FP8 block-wise (E4M3, 128x128… AIME 2025 0-shot (lighteval via l… 0.929 → 0.929 100.0%
FP8 block-wise (E4M3, 128x128… MATH-500 0-shot (lighteval via li… 0.848 → 0.848 100.0%
FP8 block-wise (E4M3, 128x128… MMLU-Pro 0-shot (lm-evaluation-ha… 0.853 → 0.853 100.0%
FP8 block-wise (E4M3, 128x128… IFEval 0-shot (lm-evaluation-harn… 0.931 → 0.927 99.6%
FP8 block-wise (E4M3, 128x128… LiveCodeBench v6 code generation … 0.773 → 0.766 99.0%
FP8 W8A8 dynamic (E4M3; weigh… LiveCodeBench v6 code generation … 0.773 → 0.789 102.0%
FP8 W8A8 dynamic (E4M3; weigh… GPQA Diamond 0-shot (lighteval vi… 0.845 → 0.849 100.4%
FP8 W8A8 dynamic (E4M3; weigh… MMLU-Pro 0-shot (lm-evaluation-ha… 0.853 → 0.854 100.0%
FP8 W8A8 dynamic (E4M3; weigh… MATH-500 0-shot (lighteval via li… 0.848 → 0.847 99.9%
FP8 W8A8 dynamic (E4M3; weigh… GSM8K-Platinum 0-shot (lm-evaluat… 0.957 → 0.956 99.9%
FP8 W8A8 dynamic (E4M3; weigh… IFEval 0-shot (lm-evaluation-harn… 0.931 → 0.927 99.6%
FP8 W8A8 dynamic (E4M3; weigh… AIME 2025 0-shot (lighteval via l… 0.929 → 0.925 99.5%
NVFP4 W4A4 (FP4 E2M1 weights … GSM8K-Platinum 0-shot (lm-evaluat… 0.957 → 0.961 100.4%
NVFP4 W4A4 (FP4 E2M1 weights … MATH-500 0-shot (lighteval via li… 0.848 → 0.85 100.2%
NVFP4 W4A4 (FP4 E2M1 weights … GPQA Diamond 0-shot (lighteval vi… 0.845 → 0.847 100.2%
NVFP4 W4A4 (FP4 E2M1 weights … IFEval 0-shot (lm-evaluation-harn… 0.931 → 0.924 99.3%
NVFP4 W4A4 (FP4 E2M1 weights … MMLU-Pro 0-shot (lm-evaluation-ha… 0.853 → 0.847 99.3%
NVFP4 W4A4 (FP4 E2M1 weights … AIME 2025 0-shot (lighteval via l… 0.929 → 0.912 98.2%
NVFP4 W4A4 (FP4 E2M1 weights … LiveCodeBench v6 code generation … 0.773 → 0.747 96.6%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/qwen3-6-35b-a3b) · [JSON](https://whatmodel.sh/api?format=json&all=1)