Kimi K2.8 Preview Takes Over Kimi Code's kimi-for-coding Model

A new model behind an unchanged name

Moonshot's Kimi Code changelog entry for September 11, 2026 says K2.8 Preview is fully rolled out, served under the existing kimi-for-coding model ID so that clients and third-party tools need no configuration changes. It follows Kimi K2.7 Code, Moonshot's June coding model. Moonshot describes the new model's coding and agent performance as close to Kimi K3, its flagship, with significantly more efficient thinking than K2.7 Code. [1]

K2.8 Preview uses the same thinking levels as K3, low, high and max, and defaults to max. With thinking turned off, requests for either K3 or K2.8 Preview are served by K2.8 Preview. Moonshot says a context window of up to 1M tokens is available on every membership tier, but its model table lists K2.8 Preview itself only on Plus plans and above, with image and video input. [1][2]

What the release notes leave out

Atlas interpretation: The changelog gives no benchmark scores, so close to K3 is Moonshot's description rather than a measurement anyone can check. It is also silent on weights. The same changelog says K3 and K2.7 Code were each released and open-sourced, and this entry says neither, so for now K2.8 Preview is something Kimi Code subscribers use rather than something anyone can download. Fully rolled out also implies an earlier staged rollout that the changelog does not date. [1]