Moonshot AI publishes Kimi K3 model weights on Hugging Face, delivering on its 11-day release promise

Moonshot AI has published the full weights of Kimi K3 on Hugging Face, delivering on the timeline it set 11 days ago when it first unveiled the 2.8-trillion-parameter model. The model card, now live at huggingface.co/moonshotai/Kimi-K3, states plainly: "We release the full Kimi K3 model weights under the Kimi K3 License."

What's new

When Moonshot introduced Kimi K3 on July 16, it took a staggered approach: API access went live immediately on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, while the underlying weights were held back with a promise of public release "by July 27." That promise has now been kept, and developers can self-host the model rather than access it only through Moonshot's hosted endpoints.

The published weights match the specs Moonshot detailed at launch:

• 2.8 trillion total parameters, with roughly 104 billion active per inference across a Mixture-of-Experts architecture
• Native multimodal support for text, images, and video, rather than a bolted-on vision module
• A 1-million-token context window
• Two new architectural components, Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which Moonshot says improve information flow across sequence length and model depth
• Weights quantized to MXFP4 with MXFP8 activations
• A lightweight detection setup built around DINOv2-style embeddings for tasks like long-horizon coding and agentic knowledge work

The weights are released under a custom Kimi K3 License rather than a fully permissive open-source license, so downstream use is still governed by Moonshot's own terms.

Context

Kimi K3 is Moonshot's second major model generation in under a year, following Kimi K2, and it landed just ahead of the World Artificial Intelligence Conference in Shanghai. At 2.8 trillion parameters, it is by a wide margin the largest open-weight model released to date, and Moonshot's own benchmarks put it ahead of Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.6 Sol and GPT-5.5 on several tasks, though those figures are the lab's own and await independent verification.

The release has not been without friction: on July 22, the White House accused Moonshot AI of covert distillation and the use of banned Nvidia chips in building Kimi K3, a dispute that adds a geopolitical layer to what is otherwise a straightforward open-weights milestone. Moonshot has not walked back its release plans in response.

Why it matters

Shipping the hosted product first and following with open weights once serving infrastructure and safety review are settled is a pattern several open-weight labs have adopted this year, and Moonshot's decision to hit its own 11-day deadline signals confidence in the release. For developers and enterprises weighing whether to self-host versus pay for frontier API access, a 2.8-trillion-parameter model with genuine long-horizon coding and reasoning claims — now downloadable rather than merely announced — raises the practical stakes of that decision, even as the geopolitical scrutiny around how it was built continues to play out.