| Benchmark | Qwen2.5 Instruct 72B | Qwen3.5 35B A3B |
|---|---|---|
| AA Intelligence | 10 | 29.9 |
| AA-LCR | 20.3 | 68.3 |
| AA-Omniscience | -52.2 | -48.1 |
| arc_challenge | 94.5 | 95.4 |
| Artificial Analysis Coding Index | 11.9 | 37 |
| C-Eval | 89.2 | 90.2 |
| critpt | 0 | 0.9 |
| GPQA Diamond | 49.1 | 84.5 |
| GSM8K | 95.8 | 90.1 |
| hellaswag | 84.8 | 85.6 |
| HLE | 4.2 | 47.4 |
| humaneval | 86.6 | 66.5 |
| IFBench | 36.9 | 72.5 |
| ifeval | 87.2 | 91.9 |
| longbench_v2 | 39.4 | 59 |
| MBPP | 72.6 | 70.8 |
| mmlu | 86.1 | 81.1 |
| mmlu_redux | 86.8 | 93.3 |
| MMLU-Pro | 71.6 | 85.3 |
| mmmlu | 74.8 | 85.2 |
| OmniScience Accuracy | 17.6 | 20.1 |
| OmniScience Non-Hallucination | 15.3 | 14.6 |
| OpenBookQA | 96.2 | 44.2 |
| scicode | 26.7 | 37.7 |
| SWE-bench Verified | 23.8 | 70 |
| Terminal-Bench Hard | 4.5 | 26.5 |
| winogrande | 82.3 | 79.2 |
| τ²-Bench Telecom (AA run) | 34.5 | 89.2 |
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.