| Benchmark | Gemma 4 E2B | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 7.4 | 14.4 |
| AA Intelligence | 10 | 7.7 |
| AA-LCR | 17 | 23 |
| AA-Omniscience | -23.6 | -68.4 |
| Artificial Analysis Coding Index | 8.3 | 6.7 |
| BFCL v3 | 56.4 | 67.3 |
| BLINK | 51.8 | 65.1 |
| chartqa | 43.5 | 84 |
| critpt | 0 | 0 |
| DocVQA-val | 0.7 | 94.7 |
| gdpval | 0 | 13.8 |
| GPQA Diamond | 43.4 | 64.1 |
| HallusionBench | 0.6 | 64.1 |
| HLE | 4.8 | 4.6 |
| IFBench | 38 | 36.6 |
| ifeval | 82.9 | 82.6 |
| LiveCodeBench v6 | 44 | 51.3 |
| mathvision | 52.4 | 60 |
| MathVista | 62.1 | 79.5 |
| MMLU-Pro | 60 | 73.6 |
| MMMU-Pro | 44.6 | 57 |
| MMStar | 57.9 | 73.2 |
| OCRBench | 0.7 | 81.6 |
| OmniScience Accuracy | 8.2 | 12.2 |
| OmniScience Non-Hallucination | 67.6 | 8.2 |
| RealWorldQA | 56.2 | 73.2 |
| scicode | 20.9 | 17.1 |
| Terminal-Bench Hard | 3 | 1.5 |
| τ²-Bench Telecom (AA run) | 22.2 | 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.