Kimi K2.5
Natively multimodal 1T/32B open model trained on about 15T mixed vision-text tokens, with Agent Swarm orchestrating up to 100 parallel sub-agents.
Joint text-vision pretraining plus zero-vision SFT; Agent Swarm trained with parallel-agent RL cuts latency up to 4.5x on parallelizable work. Launched alongside the open-source Kimi Code CLI. Became the base for Cursor's Composer 2.
- Date
- Tuesday, 27 January 2026
- Lab
- Moonshot AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 76.8% | company |
| BrowseComp (Agent Swarm) | 78.4% | company |
| AIME 2025 | 96.1% | company |
| Agent Swarm latency reduction | up to 4.5x vs single agent | company |
Modified MIT. Release date from HF README commit 2026-01-27 and Epoch AI (Epoch lists 2026-02-02, the arXiv date). Kimi's X post (quoted by Simon Willison) says Cursor's Composer 2 was built on K2.5.
Sources
- www.kimi.com/blog/kimi-k2-5
- arxiv.org/abs/2602.02276
- huggingface.co/moonshotai/Kimi-K2.5
- simonwillison.net/tags/ai-in-china/
- www.kimi.ai/blog/kimi-k2-5
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check