Kimi K2.5: benchmark results
Moonshot native multimodal MoE with 1T total and 32B active parameters, a 400M-parameter vision encoder, and 256K context. The publisher reports approximately 15T mixed visual/text tokens of continual pretraining on top of Kimi-K2-Base; this is additional training, not the whole lineage budget. Provider: Moonshot. Released 2026-01-27. Access: Open.
Aggregate Rating 1732 ± 2, rank #144 of 1656 rated models, from 2573 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ABRA - Hard | 46 | Hard-tier score (longitudinal new-lesion localisation, end-t | 100 |
| AGC-Bench - Problem Solving | 1.02 | JRT z-score | 100 |
| AGC-Bench - rpgbench | 1.74 | Dataset z-score | 100 |
| AGC-Bench - tinyfabulist | 2.03 | Dataset z-score | 100 |
| AeroGround - Trajectory Reconstruction Selection | 53 | Accuracy (%; Spatial Reasoning: select the aerial trajectory | 100 |
| Chronicles-OCR - Archaic Character Recognition | 27.1 | Exact-match accuracy (%) of the modern character given for a | 100 |
| Chronicles-OCR - Mature Script Classification | 77 | Accuracy (%) of classifying the image into one of the Seven | 100 |
| CryptoAnalystBench - Depth | 8.35 | Mean judge score from 1 to 10 for depth (technical detail, c | 100 |
| CryptoAnalystBench - Relevance | 9.58 | Mean judge score from 1 to 10 for relevance (directly addres | 100 |
| CryptoAnalystBench - Temporal Relevance | 8.78 | Mean judge score from 1 to 10 for temporal relevance (recent | 100 |
| DRAGON (Diagrams) - ChartQA | 15.1 | Evidence grounding on charts from ChartQA: given the diagram | 100 |
| DRAGON (Diagrams) - MapWise | 10.8 | Evidence grounding on maps from MapWise: given the diagram, | 100 |
Interactive version: theaggregate.ai/model?slug=kimi-k2-5 · How It Works · Data refreshed daily, snapshot 2026-10-20.