Moonshot AI Releases Kimi K3, the Largest Open Weight Model Yet

Moonshot AI published the complete weights for its new model, Kimi K3, on Hugging Face on 27 July 2026. The download comes as 96 shards totalling roughly 1.56TB, alongside a technical report and the training infrastructure tools the Beijing lab used to build it.

Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters, of which 104 billion are active for any given request. It ships with a 1 million token context window, large enough to hold a substantial codebase or a long document collection in a single prompt. Moonshot also introduced a new attention architecture for the model, Kimi Delta Attention, which the company says delivers roughly 2.5 times the scaling efficiency of its previous generation.

The release has moved fast. Hugging Face CEO Clem Delangue said Kimi K3 passed 4,000 likes within 30 minutes of going live, which he described as the platform's fastest release growth to date. That kind of reception matters for open weight models specifically, since it signals developers are willing to download and self-host something this large rather than default to a closed API.

Kimi K3's size puts it well beyond what most teams can run on their own hardware without serious infrastructure, so its immediate impact will land hardest on cloud providers and research labs able to serve a model of this scale. For everyone else, the more interesting number is the context window: a 1 million token limit paired with open weights gives smaller teams a genuine alternative to renting context length from a closed provider.

Moonshot has not said whether a smaller, distilled version of K3 is planned, which is usually the point at which a release like this becomes relevant to studios and agencies working with tighter compute budgets.

What makes this release worth tracking beyond its size is the pace of adoption. A model this large, released with full weights rather than through a paid API, tells other labs that the market for open, self-hostable frontier models is real and growing, not a niche next to the closed leaders.