Qwen 3 Max: benchmark results
Alibaba's Qwen 3 Max, the proprietary API-only flagship of the Qwen 3 family (September 2025). Provider: Alibaba. Released 2025-09-24. Access: API.
Aggregate Rating 1693 ± 4, rank #208 of 1656 rated models, from 690 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentNoiseBench (tau2-bench, User Noise) | 78 | Success on a stratified 25 percent sample of tau2-bench task | 100 |
| AlpsBench - Emotional Intelligence (English) | 3.68 | Emotional-intelligence score (1 to 5) of responses to Englis | 100 |
| BIRD-Python | 63.43 | LLM-validated execution accuracy (%; generated pandas code e | 100 |
| BIRD-Python - SQL | 63.1 | LLM-validated execution accuracy (%; generated SQL executed | 100 |
| CL4SE - Code Generation | 78.46 | Pass@1 (%) on the 636 hardest CL4SE LeetCode problems (the 3 | 100 |
| CL4SE - Code Generation (5-Shot) | 76.1 | Pass@1 (%) on the 636 hardest CL4SE LeetCode problems (the 3 | 100 |
| CL4SE - Code Review (5-Shot) | 76.88 | Accuracy (%) of the accept-or-reject decision on 1,916 GitHu | 100 |
| CL4SE - Code Summarization | 22.75 | ROUGE-L F1 (%) of the generated docstring against the origin | 100 |
| CL4SE - Patch Correctness Assessment | 76.72 | Accuracy (%) of classifying 2,274 plausible Defects4J patche | 100 |
| CL4SE - Patch Correctness Assessment (2-Shot) | 80.41 | Accuracy (%) of classifying 2,274 plausible Defects4J patche | 100 |
| Disentangling Language Roles in Multilingual L | 84 | JointSuccess (self-reported) | 100 |
| EIBench (Emotion Management) - Defense | -2.4 | Anchor-based outcome score x 100 (-100 to +100): the simulat | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-max · How It Works · Data refreshed daily, snapshot 2026-10-20.