Moonshot AI has fully released its K2.8 Preview model for Kimi Code. The model retains its existing identifier, kimi-for-coding, allowing users to access it through clients and third-party tools without requiring any configuration changes.
The company reports that the K2.8 Preview model offers comprehensive performance levels comparable to the K3 version. Developers have improved the thinking efficiency of the model, noting significant enhancements in coding and agent-based capabilities compared to the previous K2.7 Code iteration.
All subscription tiers now provide access to a context window of up to 1M tokens. Additionally, the new model supports adjustable thinking effort levels, matching the K3 series with three settings: low, high, and max, with the latter set as the default.
The system architecture is designed to route requests to the K2.8 Preview model whenever the thinking function is disabled for both the K3 series and the K2.8 Preview itself.
