| Benchmark | Gemma 4 26B A4B | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 26.1 | 10 |
| AA-LCR | 61.7 | 20.3 |
| AA-Omniscience | -50.8 | -52.2 |
| arc_challenge | 92.8 | 94.5 |
| Artificial Analysis Coding Index | 39.3 | 11.9 |
| C-Eval | 82.5 | 89.2 |
| critpt | 0 | 0 |
| GPQA Diamond | 82.3 | 49.1 |
| GSM8K | 77 | 95.8 |
| hellaswag | 85.3 | 84.8 |
| HLE | 19.3 | 4.2 |
| humaneval | 50 | 86.6 |
| IFBench | 72.4 | 36.9 |
| ifeval | 96.4 | 87.2 |
| LiveCodeBench | 79.8 | 55.5 |
| MBPP | 68.4 | 72.6 |
| mmlu | 77.8 | 86.1 |
| mmlu_redux | 92.7 | 86.8 |
| MMLU-Pro | 85.2 | 71.6 |
| mmmlu | 86.3 | 74.8 |
| OmniScience Accuracy | 19.1 | 17.6 |
| OmniScience Non-Hallucination | 13.6 | 15.3 |
| OpenBookQA | 48.8 | 96.2 |
| scicode | 40.3 | 26.7 |
| SWE-bench Verified | 57.4 | 23.8 |
| Terminal-Bench Hard | 25 | 4.5 |
| winogrande | 79.1 | 82.3 |
| τ²-Bench Telecom (AA run) | 43.6 | 34.5 |
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.