Kimi Code Launches K3-256k Model with 256k Context Window

New Model Release: Kimi K3-256k

Kimi Code has introduced a new model, Kimi K3-256k, which is now available for users. This model provides a 256k context window and is positioned alongside the existing Kimi K3 and K2.7 Code models. It is accessible via model ID in various clients and third-party tools.

Model Capabilities and Use Cases

The K3-256k model is designed for common tasks such as everyday question-and-answer interactions, code completion, routine feature development, and single-file or small-file edits. Within its 256k context limit, it delivers comparable results to the K3 (1M) model. However, it does not support video input.

Quota Consumption and Switching Recommendations

The K3-256k model consumes approximately half the quota of the K3 (1M) model. When switching from K3 (1M) to K3-256k, users are advised to manually compact their session context to under 256k to preserve task key points and maintain session integrity. Direct switching with video files in conversation history will fail, requiring prior compaction. Switching from K3-256k to K3 (1M) does not affect the cache if the 256k limit is approached.

Membership Plans and Access

Access to different models, context windows, and speeds is determined by membership plans. For instance, K3 and K3-256k models require a Moderato plan or higher. The full 1M context for K3 is available on Allegretto and higher tiers, while K3-256k has a fixed 256K context limit. Users may encounter 401 errors if their requested capabilities exceed their plan's entitlements.