| Benchmark | Qwen2.5 Instruct 72B | Qwen3.5 397B A17B |
|---|---|---|
| AA Intelligence | 10 | 34.3 |
| AA-LCR | 20.3 | 72.7 |
| AA-Omniscience | -52.2 | -30.8 |
| Artificial Analysis Coding Index | 11.9 | 48.2 |
| C-Eval | 89.2 | 93 |
| critpt | 0 | 1.7 |
| GPQA Diamond | 49.1 | 89.3 |
| GSM8K | 95.8 | 66.7 |
| HLE | 4.2 | 29 |
| IFBench | 36.9 | 78.8 |
| ifeval | 87.2 | 92.6 |
| longbench_v2 | 39.4 | 63.2 |
| mmlu_redux | 86.8 | 94.9 |
| MMLU-Pro | 71.6 | 87.8 |
| mmmlu | 74.8 | 88.5 |
| OmniScience Accuracy | 17.6 | 30.8 |
| OmniScience Non-Hallucination | 15.3 | 17.3 |
| scicode | 26.7 | 42 |
| SWE-bench Verified | 23.8 | 76.4 |
| Terminal-Bench Hard | 4.5 | 40.9 |
| τ²-Bench Telecom (AA run) | 34.5 | 95.6 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.