| Benchmark | Gemma 4 31B | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 14.4 | 14.4 |
| AA Intelligence | 29.7 | 7.7 |
| AA-LCR | 68.3 | 23 |
| AA-Omniscience | -47.9 | -68.4 |
| AI2D | 89 | 84.9 |
| Artificial Analysis Coding Index | 43.4 | 6.7 |
| critpt | 1.4 | 0 |
| ERQA | 57.5 | 47.3 |
| gdpval | 15.5 | 13.8 |
| GPQA Diamond | 85.7 | 64.1 |
| HallusionBench | 67.4 | 64.1 |
| HLE | 26.5 | 4.6 |
| IFBench | 75.6 | 36.6 |
| LiveCodeBench v6 | 80 | 51.3 |
| mathvision | 85.6 | 60 |
| MathVista | 79.3 | 79.5 |
| mmlu_redux | 93.7 | 86 |
| MMLU-Pro | 85.2 | 73.6 |
| MMMU-Pro | 76.9 | 57 |
| MMStar | 77.3 | 73.2 |
| OCRBench | 86.1 | 81.6 |
| OmniScience Accuracy | 20 | 12.2 |
| OmniScience Non-Hallucination | 18.1 | 8.2 |
| RealWorldQA | 72.3 | 73.2 |
| RefSpatial-Bench | 4.7 | 45.3 |
| scicode | 43.4 | 17.1 |
| supergpqa | 65.7 | 46.8 |
| Terminal-Bench Hard | 36.4 | 1.5 |
| τ²-Bench Telecom (AA run) | 65.5 | 15.5 |
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.