IT Home On October 23, Kimi’s official account released a 24-second pre-release video. In the video, an old-fashioned computer is running Kimi Code, and the model bar in the upper left corner shows “Model: Kimi K3.1”. Combined with previous clues, it all indicates that K3.1 is about to be released, and the developer community has already started discussions about it.
The earliest signs appeared in September this year. Several developers noticed the “kimi-k3-1” identifier in the backend model registry of the Moon’s Dark Side API platform, indicating that this model identifier passed interface detection and could be called. On the same day, the Kimi official open platform also announced a preview of the K3.1 or K3 11111 model, which supports a 1M tokens context window.
In the early morning of October 19, a submission titled “Kimi-K3.1 support for the K3 chat encoder” appeared in the SGLang repository. This submission added K3.1 dialogue encoding adaptation to the experimental routing component, along with 107 test cases.
According to ZhiXingRong’s organization, K3 originally accepted three thinking intensity levels: low, high, and max. For K3.1, the medium level was added on this basis, while the default route value remains max. The tokenizer file for K3.1 is identical to that of K3 at the byte level, using the same 16 special tokens. New tests cover dialogue formats, tool calls, and image input formats. This is a compatibility verification for encoding, not a performance benchmark. Orcarouter’s independent analysis further indicates that the submission directly states “moonshotai / Kimi-K3.1”, and cites a specific model repository revision number. This means that an identifiable K3.1 model entity has been established on Hugging Face, though it has not yet been made public.
When released in July, Kimi K3 supports a context of 1 million tokens. The official statement said that low and high modes will be added beyond the default max. If the SGLang code reflects the final product behavior, the reasoning strength of K3.1 will be expanded from three levels to four levels, with an intermediate level inserted between low and high.
After the release of K3, GitHub Copilot integrated Kimi K3 on August 6. Amazon AWS’s AI programming assistant Kiro announced its integration with K3 on October 15. Kimi became the first open-source model to participate in global distribution on AWS Bedrock. The choice of K3.1 to use Kimi Code’s interface as a pre-launch platform at least indicates that programming and developer workflows remain a key aspect of this update. In addition to the reasoning levels, the front-end cues of K3.1 also point to the 3D content creation entry “make3d” and the multi-agent collaboration marker “OK_COMPUTER_SWARM”, but these identifiers have not yet been officially confirmed by Moon’s Dark Side.
As of the announcement by IT Home, the official release date, parameter scale, or pricing for K3.1 have not been revealed.