Moonshot AI publishes Kimi K3 weights, the largest open release in AI history

The biggest open-weight release in AI history just landed.

On July 27, Moonshot AI published the full weights of Kimi K3 on Hugging Face. The model had been answering API calls since July 16; now the weights themselves are public, and anyone can download them.

The numbers

The scale is hard to picture:

• 2.8 trillion parameters in a mixture-of-experts architecture
• A 1 million token context window
• Around 1.4 TB of storage in 4-bit MXFP4; at 16-bit precision it would be roughly 5.6 TB

For comparison, most “large” open models to date have shipped at a fraction of this size. This is frontier scale, published in full.

Downloadable is not the same as runnable

Before anyone clears 1.4 terabytes of disk space, two practical caveats.

First, loading the weights takes a multi-GPU cluster of 80 GB cards. For most teams, the realistic route to using K3 is still a hosting provider rather than self-hosted hardware.

Second, check the license before building on it. The terms were not published ahead of the release, so anyone planning products on top of K3 should read what actually ships alongside the weights before committing.

Why the precedent matters

The release resets what “open model” means. Until now, the biggest models were open only through an API: you could call them, but never inspect, fine-tune or run them yourself. K3 puts frontier-scale weights in public hands, where researchers can study them, companies can self-host them, and the community can adapt them.

That changes the competitive conversation too. Every closed lab now has to answer a sharper version of an old question: what exactly does the API-only model offer that a downloadable frontier model does not?

Whether many teams can actually afford the GPUs to run it is almost beside the point. The precedent is set, and the next big release will be measured against it.