| Benchmark | DeepSeek-V3 | Gemma 4 26B A4B |
|---|---|---|
| AA Agentic Index | 1.6 | 32.1 |
| AA Intelligence | 15.4 | 26.1 |
| AA-LCR | 41.3 | 61.7 |
| AA-Omniscience | -37.6 | -50.8 |
| AIME 2025 | 51.3 | 90 |
| arc_challenge | 95.3 | 92.8 |
| Artificial Analysis Coding Index | 23 | 39.3 |
| C-Eval | 90.1 | 82.5 |
| critpt | 0 | 0 |
| gdpval | 0 | 25.7 |
| GPQA Diamond | 68.4 | 82.3 |
| GSM8K | 96.7 | 77 |
| hellaswag | 88.9 | 85.3 |
| HLE | 5.2 | 19.3 |
| hmmt_feb_2025 | 29.2 | 91.7 |
| humaneval | 92.1 | 50 |
| IFBench | 41 | 72.4 |
| ifeval | 87.3 | 96.4 |
| LiveCodeBench | 49.6 | 79.8 |
| LiveCodeBench v6 | 46.9 | 77.1 |
| MBPP | 75.4 | 68.4 |
| mmlu | 89.4 | 77.8 |
| mmlu_redux | 90.5 | 92.7 |
| MMLU-Pro | 81.2 | 85.2 |
| mmmlu | 79.4 | 86.3 |
| OmniScience Accuracy | 25.4 | 19.1 |
| OmniScience Non-Hallucination | 23.3 | 13.6 |
| OpenBookQA | 95.4 | 48.8 |
| scicode | 35.8 | 40.3 |
| supergpqa | 53.7 | 61.4 |
| SWE-bench Verified | 42 | 57.4 |
| TauBench V3 - Banking | 4.7 | 12 |
| Terminal-Bench 2.1 | 16.9 | 39 |
| Terminal-Bench Hard | 15.2 | 25 |
| winogrande | 84.9 | 79.1 |
| τ²-Bench | 32.5 | 68.2 |
| τ²-Bench Telecom (AA run) | 47.1 | 43.6 |
| τ³-Bench | 4.7 | 59 |
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.