| Benchmark | Gemma 4 31B | Gemma 4 E2B |
|---|---|---|
| AA Agentic Index | 14.4 | 7.4 |
| AA Intelligence | 29.7 | 10 |
| AA-LCR | 68.3 | 17 |
| AA-Omniscience | -47.9 | -23.6 |
| aime_2026 | 89.2 | 37.5 |
| Artificial Analysis Coding Index | 43.4 | 8.3 |
| BIG-Bench Extra Hard | 74.4 | 21.9 |
| Claw-Eval Avg | 48.5 | 53.1 |
| Codeforces ELO | 2150 | 633 |
| critpt | 1.4 | 0 |
| gdpval | 15.5 | 0 |
| GPQA Diamond | 85.7 | 43.4 |
| HallusionBench | 67.4 | 0.6 |
| HLE | 26.5 | 4.8 |
| IFBench | 75.6 | 38 |
| LiveCodeBench v6 | 80 | 44 |
| mathvision | 85.6 | 52.4 |
| MathVista | 79.3 | 62.1 |
| MedXPertQA MM | 61.3 | 23.5 |
| MMLU-Pro | 85.2 | 60 |
| mmmlu | 88.4 | 67.4 |
| MMMU-Pro | 76.9 | 44.6 |
| MMStar | 77.3 | 57.9 |
| mrcr_v2_8needle_128k_average | 66.4 | 19.1 |
| OCRBench | 86.1 | 0.7 |
| OmniDocBench 1.5 (average edit distance, lower is better) | 0.1 | 0.3 |
| OmniScience Accuracy | 20 | 8.2 |
| OmniScience Non-Hallucination | 18.1 | 67.6 |
| RealWorldQA | 72.3 | 56.2 |
| scicode | 43.4 | 20.9 |
| TauBench V3 - Banking | 14.8 | 4.5 |
| Terminal-Bench 2.1 | 43.4 | 0.4 |
| Terminal-Bench Hard | 36.4 | 3 |
| τ²-Bench | 76.9 | 24.5 |
| τ²-Bench Telecom (AA run) | 65.5 | 22.2 |
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.