Moonshot AI closes its Kimi K2 models to new API users ahead of year-end retirement

Moonshot AI, the Chinese developer of the Kimi models, has stopped offering its K2 series to developers who sign up for its API from October 31, 2026. A notice on the official model list of the Kimi platform says existing users are not affected by the cutoff, and that the company plans to retire every K2 series model by the end of December 2026.

Which models are affected

The K2 series still listed as current on the platform includes kimi-k2.7-code, a coding model with a 256,000-token context window, its faster kimi-k2.7-code-highspeed variant, and kimi-k2.6, which accepts image and text input and offers thinking and non-thinking modes. According to QCode, a third-party reseller that tracked Moonshot's documentation, these are the IDs covered by both the new-user cutoff and the year-end plan. QCode notes that Moonshot has given no exact retirement day or model-by-model order, so the December date should be treated as a plan.

This is the latest step in a gradual wind-down. Moonshot's model list shows that the original kimi-k2 preview and thinking IDs were discontinued on May 25, 2026, and that kimi-k2.5 followed on August 31.

The suggested replacement

The notice points developers to kimi-k3.1-pro. Moonshot describes it as the flagship of its K3.1 series: a "3T-class" open-weight model post-trained from Kimi K3, with native vision and a context window of one million tokens. A highspeed variant is said to run two to three times faster. The quickstart guide says the full open weights are due to be released by November 4.

Migration is not a simple model-name swap. On K3.1 Pro, thinking mode is always on; developers can only lower the reasoning_effort setting to reduce latency. Parameters such as temperature, top_p and the presence and frequency penalties are fixed and should be omitted from requests, and multi-turn or tool-calling flows must send back the complete assistant message rather than only its text. Access requires a top-up of at least $1.

Why it matters

For teams that built products on K2 models, the cutoff mainly affects new projects and new accounts, but the December retirement sets a hard deadline for existing integrations. Behavioural changes such as mandatory reasoning can alter latency, cost and output format, so applications will need testing rather than just a configuration change. QCode also observes that Moonshot tends to record such changes in its documentation rather than in separate announcements, which makes the docs the place to monitor.