Moonshot releases Kimi K3, a 2.8T-parameter open-weights multimodal model

The moonshotai organisation published Kimi K3 on Hugging Face. The card calls it an open-weight, native multimodal agentic model with 2.8 trillion parameters, 104 billion of them activated per token. It takes text, images and video and supports a 1-million-token context.

The design is new. It uses Kimi Delta Attention and Attention Residuals, with 69 KDA layers and 24 gated MLA layers, and a Stable LatentMoE that activates 16 of 896 experts. The card says this improves overall scaling efficiency by about 2.5 times over Kimi K2, which is the maker's own estimate. It also claims to be the first open 3T-class model.

The card says the aims are long-horizon coding, agentic knowledge work and reasoning. The code and weights are released under the Kimi K3 License, a custom licence, not MIT or Apache. A tech blog and a full report are linked from the card.

The card gives no release date; the date used is the Hugging Face repo creation date.

It is open in the sense that you can download it. Whether you can host it is another matter.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 3.5: Weights are public but the size is 2.8T; running it takes a cluster, and the licence is custom.
• Useful 4: Long-horizon coding and knowledge work are the stated aims, with text, image and video input.
• Notable 5: Over 11,000 likes and a 'first open 3T-class' claim on the card.
• Novel 4: New attention design (Kimi Delta Attention, Attention Residuals) and a 896-expert MoE.