| Benchmark | Llama 3.1 Instruct 405B | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 6.3 | 14.4 |
| AA Intelligence | 8.5 | 7.7 |
| AA-LCR | 24.7 | 23 |
| AA-Omniscience | -17.1 | -68.4 |
| AIME 2024 | 23.3 | 72.9 |
| Artificial Analysis Coding Index | 14.5 | 6.7 |
| critpt | 0 | 0 |
| gdpval | 0 | 13.8 |
| GPQA Diamond | 51.5 | 64.1 |
| HLE | 4.2 | 4.6 |
| IFBench | 39 | 36.6 |
| ifeval | 88.6 | 82.6 |
| LiveCodeBench | 30.1 | 51.3 |
| mmlu_redux | 86.2 | 86 |
| MMLU-Pro | 73.4 | 73.6 |
| OmniScience Accuracy | 23.2 | 12.2 |
| OmniScience Non-Hallucination | 49 | 8.2 |
| scicode | 29.9 | 17.1 |
| Terminal-Bench Hard | 6.8 | 1.5 |
| τ²-Bench Telecom (AA run) | 19 | 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.