Today (October 23), Moonshot AI released a 24-second teaser video through the official Kimi account: a vintage computer running Kimi Code, with the model bar in the top-left corner reading "Model: Kimi K3.1." The signal was brief, but it put "K3.1 is coming" front and center.
The foreshadowing actually began back in September. Multiple developers stumbled upon a "kimi-k3-1" identifier in the backend of Moonshot AI's API platform, indicating the model could already be called via the interface; on the same day, the official open platform also previewed K3.1, highlighting 1M tokens of context. In the early hours of October 19, a new commit was added to the SGLang repository, adapting the conversation encoding for K3.1 and including 107 test cases. According to a roundup by Zhixinxing, K3 originally had only three thinking-intensity levels—low, high, and max—while K3.1 inserts a medium level, with routing still defaulting to max; orcarouter pointed out that the commit explicitly mentions "moonshotai/Kimi-K3.1" and cites a specific revision number, meaning an identifiable model entity is already in place on Hugging Face, just not yet public.
More reasoning levels, but coding remains the main thread
If the code represents the final form, K3.1's reasoning intensity will expand from three levels to four. Previously, K3 was released in July with support for a million-Token context window, followed by integrations from GitHub Copilot and AWS's Kiro, making Kimi the first open-source model to participate in global revenue sharing on AWS Bedrock.
K3.1's teaser deliberately chose the Kimi Code interface, showing that coding and developer workflows remain the core narrative. The frontend also revealed a 3D creation entry point, "make3d," and a multi-agent marker, "OK_COMPUTER_SWARM," though neither has been officially confirmed. As of press time, Moonshot AI has still not announced K3.1's official date, parameter scale, or pricing.