Qwen 3.6 Max (Preview) (Thinking): benchmark results
Provider: Alibaba. Released 2026-04-20. Access: API.
Aggregate Rating 1754 ± 14, rank #226 of 2169 rated models, from 104 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HierSVA-B - Proof Success | 95.2 | Assertion proof success rate (%, C2-P): share of generated a | 100 |
| HierSVA-B - Vacuity Rate | 0.1 | Vacuous proof rate (%, C2-V): share of generated assertions | 100 |
| Nejumi 4 - jaster (0-shot) - JSeM | 85 | Exact match (%) | 100 |
| Nejumi 4 - BFCL - Live AST | 76.85 | Accuracy (%) | 99 |
| Nejumi 4 - Toxicity - Prohibited Acts | 99.48 | Criteria met (%) | 99 |
| Nejumi 4 - jaster (2-shot) - MMLU-ProX (Japanese) | 91 | Exact match (%) | 97.2 |
| Nejumi 4 - M-IFEval (Japanese) | 91.59 | Score (%) | 96.9 |
| Nejumi 4 - MT-Bench (Japanese) - Roleplay | 99.5 | Judge rating (1-10, x10) | 96.9 |
| Nejumi 4 - MT-Bench (Japanese) - Writing | 99 | Judge rating (1-10, x10) | 95.8 |
| SuperCLUE General (May 2026) - Agentic Task Planning | 83.41 | Score | 95.7 |
| Nejumi 4 - Toxicity - Fairness | 99.44 | Criteria met (%) | 94.1 |
| Nejumi 4 - jaster (2-shot) - JMMLU | 95 | Exact match (%) | 91.6 |
Interactive version: theaggregate.ai/model?slug=qwen-3-6-max-preview-thinking · How It Works · Data refreshed daily, snapshot 2026-10-23.