| Benchmark | DeepSeek-V3 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 1.6 | 14.4 |
| AA Intelligence | 15.4 | 7.7 |
| AA-LCR | 41.3 | 23 |
| AA-Omniscience | -37.6 | -68.4 |
| AIME 2024 | 59.4 | 72.9 |
| AIME 2025 | 51.3 | 74.5 |
| Artificial Analysis Coding Index | 23 | 6.7 |
| critpt | 0 | 0 |
| gdpval | 0 | 13.8 |
| GPQA Diamond | 68.4 | 64.1 |
| HLE | 5.2 | 4.6 |
| HMMT 2025 | 27.5 | 53.1 |
| IFBench | 41 | 36.6 |
| ifeval | 87.3 | 82.6 |
| LiveCodeBench | 49.6 | 51.3 |
| LiveCodeBench v6 | 46.9 | 51.3 |
| mmlu_prox | 70.5 | 65 |
| mmlu_redux | 90.5 | 86 |
| MMLU-Pro | 81.2 | 73.6 |
| OmniScience Accuracy | 25.4 | 12.2 |
| OmniScience Non-Hallucination | 23.3 | 8.2 |
| scicode | 35.8 | 17.1 |
| supergpqa | 53.7 | 46.8 |
| Terminal-Bench Hard | 15.2 | 1.5 |
| τ²-Bench Telecom (AA run) | 47.1 | 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.