- SWE-bench Verified (internal scaffold) 73.4
- SWE-bench Multilingual 67.2
- SWE-bench Pro (Qwen-corrected tasks) 49.5
- Terminal-Bench 2.0 (Terminus-2, avg of 5) 51.5
- Claw-Eval Avg 68.7
- Claw-Eval Pass^3 50
- SkillsBench Avg5 28.7
- QwenClawBench 52.6
- NL2Repo 29.4
- QwenWebBench 1397
- TAU3-Bench (gpt-5.2 user model) 67.2
- VITA-Bench (claude-4-sonnet judge) 35.6
- DeepPlanning 25.9
- Tool Decathlon 26.9
- MCPMark 37
- MCP-Atlas (public set) 62.8
- WideSearch 60.1
- MMLU-Pro 85.2
- MMLU-Redux 93.3
- SuperGPQA 64.7
- C-Eval 90
- GPQA 86
- HLE 21.4
- LiveCodeBench v6 80.4
- HMMT Feb 25 90.7
- HMMT Nov 25 89.1
- HMMT Feb 26 83.6
- IMOAnswerBench 78.9
- AIME26 (full AIME 2026 I and II) 92.7
- MMMU 81.7
- MMMU-Pro 75.3
- Mathvista(mini) 86.4
- ZEROBench_sub 34.4
- RealWorldQA 85.3
- MMBench EN-DEV-v1.1 92.8
- SimpleVQA 58.9
- HallusionBench 69.8
- OmniDocBench1.5 89.9
- CharXiv(RQ) 78
- CC-OCR 81.9
- AI2D_TEST 92.7
- RefCOCO(avg) 92
- ODInW13 50.8
- EmbSpatialBench 84.3
- RefSpatialBench 64.3
- Best of VideoMME(w sub.) 86.6 / VideoMME(w/o sub.) 82.5 86.6
- VideoMMMU 83.7
- MLVU 86.2
- MVBench 74.6
- LVBench 71.4
Qwen 3.6 35B-A3B
Qwen 3.6 35B-A3B places in the top three on 2 benchmarks.
81
Benchmarks with a score
0
Ranked first, out of five or more models
2
In the top three
Qwen Apr 2026 Qwen 3.6 35B-A3B
Open weights
144.0
#101 of 256
Epoch Capabilities Index
Self-Reported Index 144.1 #139 of 404
- MMMU-Pro 75.3% #39 /90
- Terminal-Bench 2.0 24.6% #37 /43
- NL2Repo-Bench 29.4% #27 /30
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMLU-Pro | 85.2% | #25 of 167 | Lab model card | |
| HallusionBench | 69.8% | #5 of 31 | Lab model card | |
| SEA-HELM | 73.9 | #20 of 71 | Official leaderboard | |
| MMLU-Redux | 93.3% | #14 of 45 | Lab model card | |
| SuperGPQA | 64.7% | #15 of 35 | Lab model card | |
| C-Eval | 90% | #20 of 46 | Lab model card | |
| QIMMA | 55.4% | #11 of 25 | Official leaderboard | |
| SimpleVQA | 58.9% | #20 of 30 | Lab model card | |
| Humanity's Last Exam | 21.4% | #103 of 154 | Lab model card | |
| Extended NYT Connections | 41.6% | #74 of 103 | Official leaderboard | |
| Judgemark v4 | 32.6 | #32 of 40 | Official leaderboard | |
| WideSearch | 60.1% | #18 of 19 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 84.8% | #82 of 229 | Epoch AI run | |
| OTIS Mock AIME 2024-2025 | 86.7% | #81 of 212 | Epoch AI run | |
| HMMT 2025 | 90.7% | #22 of 54 | Lab model card | |
| AIME 2026 | 92.7% | #21 of 38 | Lab model card | |
| DTBench | 73.9% | #109 of 180 | Official leaderboard | |
| HMMT Nov 2025 | 89.1% | #18 of 28 | Lab model card | |
| IMO-AnswerBench | 78.9% | #26 of 37 | Lab model card | |
| Surface Evolver Bench | 44.4% | #19 of 25 | Official leaderboard | |
| HMMT 2026 | 83.6% | #15 of 19 | Lab model card | |
| FrontierMath Tiers 1-3 v2 | 20.4% | #74 of 86 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v6 | 80.4% | #31 of 90 | Lab model card | |
| SWE-bench Verified | 73.4% | #57 of 155 | Lab model card | |
| LoopArena | 33.8% | #3 of 6 | Official leaderboard | |
| SWE-bench Multilingual | 67.2% | #50 of 73 | Lab model card | |
| SWE-Bench Pro | 49.5% | #70 of 81 | Lab model card | |
| NL2Repo-Bench | 29.4% | #27 of 30 | Lab model card | |
| SWE-bench Science | 16.8% | #9 of 10 | Official leaderboard | |
| WebDev Arena | 1,250 | #95 of 105 | arena.ai, via Epoch AI | |
| SWE-rebench | 24.7% | #12 of 13 | Official leaderboard | |
| AndroidBench | 37% | #32 of 32 | Official leaderboard | |
| QwenWebBench | 1,397 | #4 of 4 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GAIA | 56.1% | #30 of 64 | Official leaderboard | |
| SnakeBench | 24 | #94 of 188 | Official leaderboard | |
| GameWorld | 40.6% | #11 of 22 | Official leaderboard | |
| VITA-Bench | 35.6% | #8 of 16 | Lab model card | |
| DeepPlanning | 25.9% | #5 of 10 | Lab model card | |
| Mystery Game Puzzles | 22% | #45 of 84 | Epoch AI run | |
| Claw-Eval | 68.7% | #7 of 13 | Lab model card | |
| tau3-bench | 67.2% | #9 of 14 | Lab model card | |
| ClawProBench | 56.9% | #34 of 49 | Official leaderboard | |
| Toolathlon | 26.9% | #35 of 50 | Official leaderboard | |
| MCP Atlas | 62.8% | #34 of 46 | Lab model card | |
| WebHarbor | 57.4% | #3 of 4 | Official leaderboard | |
| Terminal-Bench 2.0 | 24.6% | #37 of 43 | Official leaderboard | |
| GameHorizon | 57.5 | #33 of 37 | Official leaderboard | |
| o11y-bench | 43.8% | #20 of 22 | Official leaderboard | |
| SkillsBench | 28.7% | #14 of 14 | Lab model card | |
| Claw-Eval Pass^3 | 50% | #13 of 13 | Lab model card | |
| Bench on the Clocktower | 42% | #10 of 10 | Official leaderboard | |
| MCPMark | 37% | #10 of 10 | Lab model card | |
| QwenClawBench | 52.6% | #6 of 6 | Lab model card | |
| VAKRA | 14.9% | #4 of 4 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FoodTruck Bench | $15,317 | #36 of 59 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMMU | 81.7% | #14 of 103 | Lab model card | |
| MathVista | 86.4% | #10 of 70 | Lab model card | |
| MMBench-EN v1.1 | 92.8% | #4 of 23 | Lab model card | |
| ODinW-13 | 50.8% | #3 of 13 | Lab model card | |
| RealWorldQA | 85.3% | #9 of 38 | Lab model card | |
| RefCOCO / RefCOCO+ / RefCOCOg | 92% | #7 of 29 | Lab model card | |
| GSM8K-V | 31.7% | #6 of 24 | Official leaderboard | |
| CC-OCR | 81.9% | #4 of 14 | Lab model card | |
| OCRBench v2 | 65.5% | #4 of 14 | Official leaderboard | |
| AI2D | 92.7% | #15 of 52 | Lab model card | |
| ZeroBench | 34.4% | #10 of 31 | Lab model card | |
| Video-MME | 86.6% | #17 of 51 | Lab model card | |
| MLVU | 86.2% | #8 of 22 | Lab model card | |
| EmbSpatialBench | 84.3% | #6 of 16 | Lab model card | |
| Video-MMMU | 83.7% | #14 of 36 | Lab model card | |
| MMMU-Pro | 75.3% | #39 of 90 | Lab model card | |
| RefSpatial-Bench | 64.3% | #7 of 16 | Lab model card | |
| MVBench | 74.6% | #12 of 23 | Lab model card | |
| CharXiv | 78% | #42 of 75 | Lab model card | |
| LVBench | 71.4% | #18 of 31 | Lab model card | |
| ParseBench | 37.4% | #39 of 44 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| nocot-bench (NCRI) | 61.6 | #134 of 185 | Official leaderboard |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LMCA | 29.7% | #89 of 147 | Official leaderboard | |
| Conceptual Reasoning Index | 39.7 | #101 of 151 | Official leaderboard | |
| ACCoRD | 53.0 | #110 of 152 | Official leaderboard | |
| Political Science LLM Benchmark | 64.3% | #34 of 42 | Official leaderboard |
Its lab's numbers put it at 144.1, in line with its ECI
The Self-Reported Index fits one score to the 46 capability results Qwen reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 142.9 to 145.3. How we compute it.
- MMLU-Redux 93.3
- MMBench-EN v1.1 92.8
- AIME 2026 92.7
- AI2D 92.7
- RefCOCO / RefCOCO+ / RefCOCOg 92
- HMMT 2025 90.7
- C-Eval 90
- HMMT Nov 2025 89.1
- Video-MME 86.6
- MathVista 86.4
- MLVU 86.2
- GPQA Diamond 86
- RealWorldQA 85.3
- MMLU-Pro 85.2
- EmbSpatialBench 84.3
- Video-MMMU 83.7
- HMMT 2026 83.6
- CC-OCR 81.9
- MMMU 81.7
- LiveCodeBench v6 80.4
- IMO-AnswerBench 78.9
- CharXiv 78
- MMMU-Pro 75.3
- MVBench 74.6
- SWE-bench Verified 73.4
- LVBench 71.4
- HallusionBench 69.8
- Claw-Eval 68.7
- SWE-bench Multilingual 67.2
- tau3-bench 67.2
- SuperGPQA 64.7
- MCP Atlas 62.8
- WideSearch 60.1
- SimpleVQA 58.9
- Terminal-Bench 2.0 51.5
- ODinW-13 50.8
- Claw-Eval Pass^3 50
- SWE-Bench Pro 49.5
- MCPMark 37
- VITA-Bench 35.6
- ZeroBench 34.4
- NL2Repo-Bench 29.4
- SkillsBench 28.7
- Toolathlon 26.9
- DeepPlanning 25.9
- Humanity's Last Exam 21.4
Its lab's numbers next to 3 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap | What differs |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 51.5% Terminal-Bench 2.0 (Terminus-2, avg of 5) | 24.6% Leaderboard, little-coder | +27 pts | Qwen uses Terminus-2 with a three-hour timeout, while the compared leaderboard entry uses little-coder. |
| GPQA Diamond | 86% GPQA | 84.8% Epoch AI run, none | +1.2 pts | |
| Toolathlon | 26.9% Tool Decathlon | 26.9% Leaderboard | Same |
Qwen's models over time
Qwen's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.