Qwen3.8 Max
Qwen3.8 MaxCategory scores
Scores per capability area with the model's rank among ranked models in that category.
- 1Instruction following#294.7
- 2Multimodal#491.2
- 3Reasoning87.0
- 4Agentic & tool use#1482.9
- 5Math78.9
- 6Knowledge#2571.9
- 7Coding#3670.6
Benchmark results (59)
As reported in system cards, papers and public evaluations. Position compares this model with every other model reporting the same benchmark.
Multimodal
| MMMU-Pro | 82.3 | #4 of 48 |
| CharXiv | 93.5 | #1 of 42 |
| MathVision | 95.2 | #1 of 20 |
| SimpleVQA | 75.0 | #3 of 17 |
| RealWorldQA | 88.0 | #2 of 16 |
| ERQA | 77.8 | #1 of 16 |
| ScreenSpot Pro | 84.5 | #4 of 15 |
| CharXiv w/o tools | 88.4 | #2 of 14 |
| VideoMMMU | 88.7 | #1 of 12 |
| OmniDocBench 1.5 | 92.1 | #1 of 11 |
| Video-MME (with subtitle) | 90.4 | #1 of 11 |
| MLVU (M-Avg) | 90.8 | #1 of 9 |
| BabyVision | 82.0 | #1 of 8 |
| ZeroBench | 24.0 | #5 of 8 |
| MedXpertQA (MM) | 80.4 | #2 of 8 |
| CC-OCR | 79.6 | #5 of 7 |
| MMVU | 82.4 | #1 of 7 |
| LVBench | 81.8 | #4 of 6 |
| MathVision w/ Python | 97.7 | #2 of 5 |
| BabyVision w/ Python | 91.3 | #2 of 5 |
| OCRBench V2 | 74.2 | #1 of 5 |
| Vision2Web | 69.0 | #1 of 4 |
| ZeroBench w/ Python | 49.0 | #2 of 3 |
| PerceptionBench | 63.5 | #1 of 3 |
Agentic & tool use
| Terminal-Bench 2.1 | 86.6 | #13 of 68 |
| MazeBench | 0 | #20 of 42 |
| OSWorld-Verified | 86.1 | #1 of 35 |
| JobBench | 53.4 | #12 of 34 |
| HLE w/ tools | 56.2 | #11 of 27 |
| AutomationBench | 27.3 | #23 of 25 |
| OSWorld 2.0 | 19.4 | #15 of 24 |
| Toolathlon-Verified | 72.5 | #15 of 22 |
| Agents' Last Exam | 52.4 | #3 of 17 |
| WideResearch | 81.9 | #3 of 16 |
| Hermes Index | 36.2 | #11 of 16 |
| Hermes Bench | 70.5 | #10 of 16 |
| AndroidWorld | 85.3 | #2 of 8 |
| CoWorkBench | 74.8 | #2 of 4 |
| WebArena-Verified | 66.8 | #2 of 4 |
| MobileWorld | 77.8 | #1 of 1 |
Coding
| SWE-bench Pro | 67.7 | #9 of 69 |
| DeepSWE | 56.6 | #40 of 50 |
| NL2Repo | 55.9 | #8 of 31 |
| Bug Hunt Bench | 25.7 | #15 of 23 |
| FrontierSWE v2 | 15.8 | #18 of 20 |
| VulcanBench v3 | 81.2 | #11 of 14 |
| OpenHarmony Bench | 60.8 | #1 of 12 |
| MLS-Bench Lite | 41.0 | #2 of 4 |
| FrontierSWE | 73.5 | #3 of 3 |
| PaperBench | 93.0 | #1 of 1 |
| QwenReactBench | 1724 | #1 of 1 |
Knowledge
| GPQA | 92.6 | #11 of 117 |
| HLE | 43.6 | #25 of 69 |
| HLE w/o tools | 43.6 | #13 of 43 |
Reasoning
| LongBench v2 | 66.3 | #2 of 17 |
| MRCRv2 | 92.9 | #1 of 12 |
Math
| FrontierMath v2 (Tiers 1-3) | 74.737 | #20 of 64 |
| FrontierMath v2 (Tier 4) | 46.341 | #19 of 51 |
Instruction following
| IFBench | 82.8 | #2 of 42 |
Pricing by source
USD per 1M tokens.
| Source | Input | Output | Cached in |
|---|---|---|---|
| List price | — | — | — |
Qwen publishes the Qwen3.8-2.4T-A95B checkpoint for self-hosting under the custom Qwen3.8-Max License. Alibaba Cloud Model Studio's official pricing page separately lists the hosted qwen3.8-max SKU in both non-thinking and thinking modes for the 0 < Token <= 1M tier. The China and global tables list CNY 12 input / CNY 36 output per million tokens; the US international table lists CNY 14.988 input / CNY 44.965 output. The USD numeric fields stay null because we do not convert a non-USD first-party price.
Arena ratings
Human-preference ratings from LMArena. Best-rated variant shown.
| Text | 1483 | #17 · 27.8K votes |
| WebDev (coding)(0902) | 1674 | #11 · 14.4K votes |
| Vision | 1313 | #10 · 10K votes |
| Search | 1261 | #2 · 1.9K votes |
Speed
Measured by us through OpenRouter (median of trials).
Not available on OpenRouter, so we cannot measure it yet.
Related models
Same family
Similar score
- Step 5 Preview69.6 · $1.43
- Kimi K370.0 · $6.00
- Claude Opus 4.870.0 · $10.00
- GPT-5.6 Terra70.2 · $4.50
- GPT-5.569.2 · $11.25