Reportedly: Moonshot AI teases Kimi K3.1 model in preview video
Model: Kimi K3 · availability, license and releases
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Moonshot AI released a 24-second preview video showing a computer running Kimi Code with 'Model: Kimi K3.1' displayed, IT Home reported. The video was published on October 23, 2026, and no official release date has been announced.
China context
- Original name
- 月之暗面
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China should monitor the Hugging Face repository 'moonshotai/Kimi-K3.1' for weight availability, which would enable local deployment.
- For investors
- The teaser and repository references suggest Moonshot AI is preparing a new model release, which could impact the competitive landscape for open-weights models.
Translated from Chinese. Quotes and facts link to the original sources.
Moonshot AI published a 24-second teaser video on its official Kimi account showing an old computer running Kimi Code with the model selector displaying 'Model: Kimi K3.1'. The video was reported by IT Home on October 23, 2026. Earlier signs of K3.1 included a 'kimi-k3-1' identifier appearing in the Moonshot AI API platform's backend model registry in September 2026, and a preview on the Kimi open platform indicating support for a 1M token context window. On October 19, 2026, a commit titled 'Kimi-K3.1 support for the K3 chat encoder' appeared in the SGLang repository, adding chat encoding adaptation for K3.1 with 107 test cases. According to an analysis by a Chinese AI community, K3.1 adds a 'medium' thinking intensity level to the existing low, high, and max options, with the routing default remaining max. The tokenizer files are byte-identical to K3, using the same 16 special tokens, and new tests cover chat format, tool calling, and image input format, indicating encoding compatibility rather than performance benchmarking. An independent analysis by orcarouter noted that the commit explicitly references 'moonshotai / Kimi-K3.1' and a specific model repository revision, suggesting a recognizable K3.1 model entity may already exist on Hugging Face.
The SGLang commit adds experimental routing support for the K3.1 chat encoder, with 107 test cases covering chat format, tool calling, and image input. The tokenizer is byte-identical to K3, and K3.1 introduces a 'medium' thinking intensity level while keeping the default at 'max'.
Developers using Moonshot AI's API will need to adapt to a new 'medium' thinking intensity option and potential changes in model routing when K3.1 becomes available. The appearance of a Hugging Face repository reference suggests the model weights may be released, which would allow outside-China developers to run it locally.
If K3.1 is released with a 1M token context window and a new thinking intensity level, it could improve performance for long-context tasks and offer more control over reasoning depth, potentially affecting API costs for developers.
The next verifiable signal is an official announcement from Moonshot AI with a release date and model specifications. Another signal is the public availability of the 'moonshotai/Kimi-K3.1' repository on Hugging Face, which would confirm open weights.