| Benchmark | Kimi K2.5 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 52.8 | 14.4 |
| AA Intelligence | 36 | 7.7 |
| AA-LCR | 73 | 23 |
| AA-Omniscience | -7.3 | -68.4 |
| AIME 2025 | 96.1 | 74.5 |
| Artificial Analysis Coding Index | 46.8 | 6.7 |
| critpt | 3.1 | 0 |
| gdpval | 38.3 | 13.8 |
| GPQA Diamond | 87.9 | 64.1 |
| HLE | 50.2 | 4.6 |
| HMMT 2025 | 95.4 | 53.1 |
| IFBench | 70.2 | 36.6 |
| InfoVQA (val) | 92.6 | 79.5 |
| LiveCodeBench v6 | 85 | 51.3 |
| LongVideoBench | 79.8 | 57.7 |
| LVBench | 75.9 | 53.5 |
| mathvision | 84.2 | 60 |
| MathVista | 90.1 | 79.5 |
| MMLU-Pro | 87.1 | 73.6 |
| MMMU-Pro | 78.5 | 57 |
| OCRBench | 92.3 | 81.6 |
| OmniScience Accuracy | 35.2 | 12.2 |
| OmniScience Non-Hallucination | 50 | 8.2 |
| OSWorld-Verified | 63.3 | 31.4 |
| scicode | 49 | 17.1 |
| Terminal-Bench Hard | 34.8 | 1.5 |
| Video-MME | 87.4 | 59.7 |
| VideoMMMU | 86.6 | 69.4 |
| τ²-Bench Telecom (AA run) | 95.9 | 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.