Moonshot AI said in a Kimi Code changelog on Sept. 11 that K2.8 Preview is now live for all users. The company describes it in two lines: overall performance close to K3, with thinking efficiency clearly improved over K2.7 Code, and "comprehensive gains" in both coding and agent capabilities.
This swap didn't come with a new name. The model ID is still kimi-for-coding, so the Kimi Code client and any third-party tool wired to that ID need no configuration change — requests land on the new model automatically.
Three confirmed changes
First, thinking tiers. K2.8 Preview supports adjustable thinking effort, matching K3's tiers of low, high, and max, with max as the default.
Second, context length. Every membership tier now gets a 1M-token context window, with no more tier-based limits.
Third, routing. The changelog states that once thinking is turned off, requests to both the K3 series and K2.8 Preview are routed to K2.8 Preview without thinking. In practice, a user who selects K3 in Kimi Code but disables thinking actually gets K2.8 Preview.
The claim of being "close to K3" isn't backed by numbers yet. The changelog includes no benchmark table, no date for when the preview becomes a stable release, and no mention of whether membership pricing or quotas will change. For comparison, when K2.7 Code launched on June 12, Moonshot AI cited two figures: a 10.4% improvement on Program-Bench and a 30% drop in reasoning token usage. K3 itself launched on July 16, with 2.8 trillion total parameters and a 1 million-token context window.
Same model name, different stakes for different users
kimi-for-coding functions as a fixed alias — which model sits behind it is Moonshot AI's call. For individual developers, that removes the need to touch any settings.
For teams that have wired Kimi Code into CI pipelines, automation scripts, or a shared coding tool, the situation is different: the same configuration now runs against a different model as of today, and output style, latency, and token usage could all shift. The changelog doesn't say whether K2.7 Code can still be specified directly.
Competition among domestic coding subscriptions has gotten crowded. Zhipu AI kicked off a citywide "Coding" program in Hangzhou on Sept. 10, offering subsidies for its Coding plans. Kimi's changelog this time carries no pricing line — what it's leading with is the new model and the 1M-token context across every tier.
Permission mode changed twice in two days
Separate from the model swap, the Kimi Code command-line client revised its permission logic twice in early September. The Sept. 2 release, v0.40.0, renamed the old YOLO and Auto tiers to "fully automatic" (/never-ask) and "ask when needed" (/ask-when-needed), with Manual renamed to "always ask." It also added a built-in guardrail: commands such as shutdown, reboot, and rm -rf were rejected outright in fully automatic mode, and required user confirmation before running in the other modes.
Two days later, v0.41.0 stated: "Dangerous commands are no longer blocked. Auto-permission mode no longer blocks dangerous commands or commands that can't be statically analyzed." The documentation doesn't explain why the block was reversed, nor whether "auto-permission mode" here refers only to "fully automatic" or also covers "ask when needed."
v0.41.0 also added two web features: an experimental multi-agent collaboration mode called Tower (launched with /tower), and the ability to highlight text and leave comments in messages, file previews, and diffs. In the release before that, the sub-agent model pool moved from an experimental feature to a stable one, letting users configure a set of candidate models.
For teams that run Kimi Code in fully automatic mode inside unattended environments, whether delete and shutdown commands now get blocked depends on the client version actually installed — it needs to be tested directly. Moonshot AI hasn't issued a separate statement on this change.
Sources: Kimi Code official changelog (K2.8 Preview, v0.40.0, v0.41.0 entries), IT Home, CocoLoop; figures for K3 and K2.7 Code follow Moonshot AI's official statements.