- Agentic coding DeepSWE 1.1 (best of Claude Code and mini-SWE-agent) 58.7
- Agentic coding SWE-bench Pro (Claude Code, Qwen-corrected tasks) 62.5
- SWE-bench Multilingual (mini-SWE-agent) 81
- Repo-level code generation NL2Repo-Bench (Claude Code) 48.1
- Long-horizon office work CoWorkBench 73.9
- Professional job tasks JobBench 55.7
- Agents' Last Exam Pass@1 24.3 Score 51.2 24.3
- Real-world tool use Toolathlon Verified (Pass@1) 73.5
- Instruction following IFBench 81.3
- Scientific reasoning GPQA Diamond 91.7
- Multidisciplinary reasoning HLE (judged by GPT-4o) 35.9
- Competitive coding LiveCodeBench v6 91.9
- Multimodal tool use ClawEval-MM Pass@3 64.4 Average 60.4 60.4
- Application recreation RecreationBench 49.9
- Mobile use AndroidWorld 84.5
- Computer use OSWorld 2.0 Binary 19.4 Partial 52.3 19.4
- Visual web development Vision2Web 64
- Embodied intelligence ERQA 72.3
- Long video understanding LVBench 76.6
- Real-world perception RealWorldQA 88.5
- Visual math problem solving MathVision Without CI 90.6 With CI 95.7 95.7
- Scientific chart analysis CharXiv (RQ) Without CI 84.6 With CI 90.6 90.6
Qwen3.8 Flash
Qwen3.8 Flash ranks first on RealWorldQA.
Qwen Released 26 Aug 2026 27
Benchmarks with a score
1
Ranked first, out of five or more models
8
In the top three
Qwen Aug 2026 Qwen3.8 Flash
Closed weights
No Epoch Capabilities Index score yet
- IFBench 81.3% #6 /70
- NL2Repo-Bench 48.1% #13 /28
- ERQA 72.3% #3 /26
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 | |
|---|---|---|---|---|
| IFBench | 81.3% | #6 of 70 | Model card | |
| AgentIF (2026-10 setting) | 65.6% | #3 of 24 | Official leaderboard | |
| Humanity's Last Exam | 35.9% | #66 of 147 | Model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 91.7% | #18 of 284 | Model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v6 | 91.9% | #4 of 76 | Model card | |
| SWE-bench Multilingual | 81% | #11 of 70 | Model card | |
| SWE-Bench Pro | 62.5% | #21 of 77 | Model card | |
| NL2Repo-Bench | 48.1% | #13 of 28 | Model card | |
| Vision2Web | 64% | #3 of 5 | Model card | |
| DeepSWE | 58.7% | #38 of 51 | Model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Toolathlon | 73.5% | #6 of 37 | Model card | |
| AndroidWorld | 84.5% | #3 of 18 | Model card | |
| RecreationBench | 49.9% | #2 of 3 | Model card | |
| OSWorld 2.0 | 19.4% | #19 of 19 | Model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CoWorkBench | 73.9% | #3 of 6 | Model card | |
| JobBench | 55.7% | #8 of 12 | Model card | |
| Agents' Last Exam | 24.3% | #22 of 25 | Model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| RealWorldQA | 88.5% | #1 of 30 | Model card | |
| MATH-Vision (MathVision) | 95.7% | #3 of 42 | Model card | |
| CharXiv | 90.6% | #6 of 67 | Model card | |
| ERQA | 72.3% | #3 of 26 | Model card | |
| LVBench | 76.6% | #8 of 25 | Model card | |
| Claw-Eval-MM | 60.4% | #3 of 7 | Model card | |
| Humanity's Sixth Sense | 30.9% | #15 of 29 | Official leaderboard | |
| ScienceVision | 0.36 | #7 of 11 | Official leaderboard | |
| GDP.pdf | 16.6% | #28 of 39 | Epoch AI |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SuperCLUE | 66.4 | #8 of 18 | Official leaderboard |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.