Across 12 shared benchmarks, Gemma 3 27B scores higher on 4 and Qwen2.5 VL 32B Instruct on 7, with 1 level. The widest gap is mathvision, where Gemma 3 27B scores 46 against 40.
| Benchmark | Gemma 3 27B | Qwen2.5 VL 32B Instruct |
|---|---|---|
| arena_vision | 1165 | 1153 |
| docvqa | 86.6 | 94.8 |
| GPQA Diamond | 46 | 46 |
| humaneval | 87.8 | 91.5 |
| MATH | 89 | 82.2 |
| mathvision | 46 | 40 |
| MathVista | 67.6 | 74.7 |
| MBPP | 74.4 | 84 |
| mmlu | 76.9 | 78.4 |
| MMLU-Pro | 67.6 | 68.8 |
| MMMU | 64.9 | 70 |
| MMMU-Pro | 49.7 | 49.5 |
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.