| Benchmark | Gemma 4 E2B | Kimi K2.5 |
|---|---|---|
| AA Agentic Index | 7.4 | 52.8 |
| AA Intelligence | 10 | 36 |
| AA-LCR | 17 | 73 |
| AA-Omniscience | -23.6 | -7.3 |
| aime_2026 | 37.5 | 95.8 |
| Artificial Analysis Coding Index | 8.3 | 46.8 |
| critpt | 0 | 3.1 |
| gdpval | 0 | 38.3 |
| GPQA Diamond | 43.4 | 87.9 |
| HLE | 4.8 | 50.2 |
| IFBench | 38 | 70.2 |
| LiveCodeBench v6 | 44 | 85 |
| mathvision | 52.4 | 84.2 |
| MathVista | 62.1 | 90.1 |
| MMLU-Pro | 60 | 87.1 |
| MMMU-Pro | 44.6 | 78.5 |
| OCRBench | 0.7 | 92.3 |
| OmniScience Accuracy | 8.2 | 35.2 |
| OmniScience Non-Hallucination | 67.6 | 50 |
| scicode | 20.9 | 49 |
| TauBench V3 - Banking | 4.5 | 14.2 |
| Terminal-Bench 2.1 | 0.4 | 45.7 |
| Terminal-Bench Hard | 3 | 34.8 |
| τ²-Bench | 24.5 | 85.4 |
| τ²-Bench Telecom (AA run) | 22.2 | 95.9 |
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.