Moonshot AI has published open weights for Kimi K3 on Hugging Face and GitHub, describing it as a native multimodal, agentic model with 2.8 trillion parameters and a 1 million-token context window. The vLLM and LMSYS projects also announced launch-day serving support, giving developers early options for deploying th...
Moonshot AI has released open weights for Kimi K3, publishing the model through official pages on Hugging Face and GitHub.
According to the official Hugging Face model card for moonshotai/Kimi-K3, Kimi K3 is an open-weight, native multimodal, agentic model with 2.8 trillion parameters and a 1 million-token context window. The same model card says the full model weights are released under the Kimi K3 License.
Moonshot AI’s official GitHub repository for Kimi K3 mirrors the release materials and describes the model as a 2.8T open-weight multimodal agentic model. The repository also says Kimi K3 is built on Kimi Delta Attention and Attention Residuals, two architectural elements Moonshot AI highlights as part of the release.
The public availability of weights is the main practical change for developers: instead of only using a hosted interface, teams can examine the released materials and work with the model through supported serving software, subject to the Kimi K3 License and their own infrastructure constraints.
Two serving ecosystems announced day-0 support for Kimi K3.
The vLLM project, in a post titled “Kimi K3 Is Here: Efficient Day-0 Support on vLLM,” says Moonshot AI’s Kimi K3 weights are public and provides serving commands and production deployment guidance for running moonshotai/Kimi-K3. vLLM frames its post around helping users deploy the model efficiently at launch.
LMSYS Org also published a post saying SGLang and Miles added launch-day support for Kimi K3. LMSYS identifies the model as a 2.8T open 3T-class model with a 1 million-token context window and native visual understanding.
Together, those posts indicate that Kimi K3 is not just a model-card release: major open serving projects moved quickly to make it runnable for users who have the required hardware and operational setup.
For developers and research teams, the key points are the combination of open weights, very large scale, long-context capacity, and multimodal support. Moonshot AI’s own materials position Kimi K3 as an agentic model, while the Hugging Face and GitHub pages emphasize native multimodality and the 1 million-token context window.
That does not mean the model will be easy or inexpensive to run. A 2.8T-parameter model is a major infrastructure undertaking, and deployment will depend on serving frameworks, hardware availability, memory requirements, and license terms. The vLLM and LMSYS posts are therefore important because they provide early implementation paths, but they do not remove the operational complexity of hosting a model at this scale.
The available official and project sources confirm that Moonshot AI released Kimi K3 weights and that vLLM, SGLang, and Miles announced early support. They do not establish claims about any specific user’s code being used to train future models, nor do they address national-security determinations by the Canadian government.
For now, the source-backed story is narrower and clearer: Moonshot AI has made Kimi K3’s weights publicly available under its license, and prominent open serving projects have added launch-day support for developers seeking to deploy it.
Moonshot AI has released open weights for Kimi K3, publishing the model through official pages on Hugging Face and GitHub.
The same model card says the full model weights are released under the Kimi K3 License.
Moonshot AI’s official GitHub repository for Kimi K3 mirrors the release materials and describes the model as a 2.8T open weight multimodal agentic model.
Continue reading
