Moonshot AI is hinting at the imminent release of its new Kimi K3.1 model. This anticipation is fueled by a brief video showcasing the name “Kimi Code” alongside “Model: Kimi K3.1” displayed on an older computer. While an official release date remains unannounced, the model’s presence within developer communities and updates to software repositories strongly suggest its arrival is near.
Early Signs and Development
The first indications of Kimi K3.1’s existence emerged in September when developers spotted “kimi-k3-1” referenced in the internal registry of Moonshot AI’s API platform. This pointed to the potential for invoking a new model via API. Around the same time, preliminary announcements for K3.1, or K3 11111, appeared on the open Kimi platform, indicating support for context windows up to 1 million tokens.
On October 19th, the SGLang repository logged a commit for “Kimi-K3.1 support for K3 chat encoder.” This change integrated chat encoder adaptations for K3.1 into an experimental routing component, complete with 107 test cases. Analysis of this commit suggests that K3.1, unlike its predecessor K3 which offered three processing modes (low, high, max), will introduce a fourth, intermediate mode named “medium.” The “max” mode is expected to remain the default.
The tokenizer files for K3.1 are byte-for-byte identical to those of K3, utilizing the same 16 special tokens. New tests encompass dialogue formatting, tool calls, and image inputs, indicating a focus on encoding compatibility rather than performance benchmarking.
Further analysis by orcarouter confirmed that the commit directly references “moonshotai / Kimi-K3.1” and specifies a particular revision number from the model’s repository. This implies that the identified K3.1 model entity already exists on the Hugging Face platform, despite not having been officially revealed.
Enhancements and Integrations
Following the July release of Kimi K, which supported a 1 million token context, developers had announced the addition of “low” and “high” modes alongside “max.” If the functionality reflected in SGLang aligns with the final product, K3.1 would expand the inference intensity options from three to four, inserting a medium level between “low” and “high.”
Following K3’s launch, GitHub Copilot integrated it on August 6th. On October 15th, Amazon AWS AI Programming Assistant Kiro announced its integration with K3, making Kimi the first open model to participate in AWS Bedrock’s revenue sharing program. The use of the Kimi Code interface for previewing K3.1 underscores that developer programming and workflows remain a key narrative for this update.
Beyond inference intensity modes, early data points to potential new features such as “make3d” for 3D content generation and “OK_COMPUTER_SWARM” for indicating multi-agent collaboration. However, these indicators have not yet been officially confirmed by Moonshot AI.
What’s Next?
As of this article’s publication, Moonshot AI has not provided official information regarding Kimi K3.1’s release date, specifications, or pricing. Nevertheless, numerous technical indicators and preview videos suggest the model is poised for launch, with developers eagerly awaiting its formal introduction.
