| Benchmark | Qwen2.5 Instruct 72B | Qwen3.5 27B |
|---|---|---|
| AA Intelligence | 10 | 34.6 |
| AA-LCR | 20.3 | 72.3 |
| AA-Omniscience | -52.2 | -44 |
| Artificial Analysis Coding Index | 11.9 | 34.9 |
| C-Eval | 89.2 | 90.5 |
| critpt | 0 | 0.9 |
| GPQA Diamond | 49.1 | 85.8 |
| HLE | 4.2 | 48.5 |
| IFBench | 36.9 | 76.5 |
| ifeval | 87.2 | 95 |
| longbench_v2 | 39.4 | 60.6 |
| mmlu_redux | 86.8 | 93.2 |
| MMLU-Pro | 71.6 | 86.1 |
| mmmlu | 74.8 | 85.9 |
| OmniScience Accuracy | 17.6 | 20.7 |
| OmniScience Non-Hallucination | 15.3 | 24.9 |
| scicode | 26.7 | 39.5 |
| SWE-bench Verified | 23.8 | 75 |
| Terminal-Bench Hard | 4.5 | 32.6 |
| τ²-Bench Telecom (AA run) | 34.5 | 93.9 |
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.