Kimi K2-Instruct-0905
K2 refresh that doubles context from 128K to 256K and improves agentic coding, with SWE-bench Verified rising from 65.8 to 69.2 and Terminal-Bench from 37.5 to 44.5.
Same 1T/32B architecture, improved post-training for coding agents and front-end work, and double the context. Moonshot's own five-run means, using an in-house harness derived from SWE-agent.
- Date
- Friday, 5 September 2025
- Lab
- Moonshot AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 69.2 ± 0.63 vs 65.8 for K2-0711; mean of 5 runs | company |
| Terminal-Bench | 44.5 ± 2.03 vs 37.5 for K2-0711 (Terminus-2) | company |
The dates conflict. Weights were uploaded to HF on 2025-09-03, SCMP on 2025-09-04 said release was imminent, and the README was finalized 2025-09-05, while Wikipedia says 2025-09-09. Used 2025-09-05.
Sources
- huggingface.co/moonshotai/Kimi-K2-Instruct-0905
- www.scmp.com/tech/tech-trends/article/3324350/moonshot-ais-updated-kimi-model-offers-expan
This record was checked against its sources on 6 October 2026. How we check