| Benchmark | Gemma 4 E2B | Qwen3.5 27B |
|---|---|---|
| AA Agentic Index | 7.4 | 54.6 |
| AA Intelligence | 10 | 34.6 |
| AA-LCR | 17 | 72.3 |
| AA-Omniscience | -23.6 | -44 |
| aime_2026 | 37.5 | 92.6 |
| Artificial Analysis Coding Index | 8.3 | 34.9 |
| bfcl_v4 | 37 | 68.5 |
| Claw-Eval Avg | 53.1 | 64.3 |
| critpt | 0 | 0.9 |
| gdpval | 0 | 33 |
| GPQA Diamond | 43.4 | 85.8 |
| HallusionBench | 0.6 | 70 |
| HLE | 4.8 | 48.5 |
| IFBench | 38 | 76.5 |
| ifeval | 82.9 | 95 |
| LiveCodeBench v6 | 44 | 80.7 |
| mathvision | 52.4 | 86 |
| MathVista | 62.1 | 87.8 |
| MMLU-Pro | 60 | 86.1 |
| mmmlu | 67.4 | 85.9 |
| MMMU-Pro | 44.6 | 75 |
| MMStar | 57.9 | 81 |
| OCRBench | 0.7 | 89.4 |
| OmniScience Accuracy | 8.2 | 20.7 |
| OmniScience Non-Hallucination | 67.6 | 24.9 |
| RealWorldQA | 56.2 | 83.7 |
| scicode | 20.9 | 39.5 |
| Terminal-Bench Hard | 3 | 32.6 |
| τ²-Bench Telecom (AA run) | 22.2 | 93.9 |
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.