Across 30 shared benchmarks, Gemma 4 E2B scores higher on 2 and Qwen3.5 122B A10B on 28. The widest gap is HLE, where Qwen3.5 122B A10B scores 47.5 against 4.8.
| Benchmark | Gemma 4 E2B | Qwen3.5 122B A10B |
|---|---|---|
| AA Agentic Index | 7.4 | 21.3 |
| AA Intelligence | 10 | 32.8 |
| AA-LCR | 17 | 70.3 |
| AA-Omniscience | -23.6 | -41.5 |
| Artificial Analysis Coding Index | 8.3 | 45.7 |
| bfcl_v4 | 37 | 72.2 |
| critpt | 0 | 0.9 |
| gdpval | 0 | 24.3 |
| GPQA Diamond | 43.4 | 86.6 |
| HallusionBench | 0.6 | 67.6 |
| HLE | 4.8 | 47.5 |
| IFBench | 38 | 76.1 |
| ifeval | 82.9 | 93.4 |
| LiveCodeBench v6 | 44 | 78.9 |
| mathvision | 52.4 | 86.2 |
| MathVista | 62.1 | 87.4 |
| MMLU-Pro | 60 | 86.7 |
| mmmlu | 67.4 | 86.7 |
| MMMU-Pro | 44.6 | 76.9 |
| MMStar | 57.9 | 82.9 |
| OCRBench | 0.7 | 92.1 |
| OmniScience Accuracy | 8.2 | 24.4 |
| OmniScience Non-Hallucination | 67.6 | 12.9 |
| RealWorldQA | 56.2 | 85.1 |
| scicode | 20.9 | 42 |
| t2-bench | 29.4 | 79.5 |
| TauBench V3 - Banking | 4.5 | 15.3 |
| Terminal-Bench 2.1 | 0.4 | 47.6 |
| Terminal-Bench Hard | 3 | 31.1 |
| τ²-Bench Telecom (AA run) | 22.2 | 93.6 |
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.