Moonshot releases Kimi K3, a 2.8-trillion-parameter open model for server-class self-hosting
Originally announced by Moonshot AI . We link the primary source so you can read it for yourself.
Moonshot AI has released the open weights for Kimi K3, a 2.8-trillion-parameter model that sits at or near the top of several independent open-weight leaderboards. It is worth covering both for what it is and for what it says about the range of what “local” means.
The news
Kimi K3 is a sparse mixture-of-experts model with 896 experts, of which 16 are active for any given token, and a context window of 1,048,576 tokens. It uses a hybrid attention design, combining linear and full attention, that Moonshot calls Kimi Delta Attention, aimed at handling that very long context efficiently. The weights were released on 27 July 2026, with support for the model landing in vLLM.
On independent evaluations it ranks highly: it placed near the top of the Artificial Analysis Intelligence Index and the Vals leaderboards, and led a preliminary Arena web-development ranking. Those are third-party leaderboards rather than vendor-reported figures, which makes them more useful than a developer’s own numbers, though leaderboards are still an imperfect guide rather than a verdict.
One caveat matters. As of shortly after launch, the licence terms for the weights were not clearly published. Moonshot’s earlier Kimi K2 used a modified MIT licence, so a permissive release is plausible, but until K3’s terms are confirmed we would not assume them, and nor should you.
Why this needs the word “server-class”
Here is the honest part. Kimi K3 is open-weight and self-hostable, but it is not a model you run on a laptop, or even on a high-end workstation. At 2.8 trillion parameters, its weights are roughly 1.4TB even in a native 4-bit format, and Moonshot recommends a configuration of 64 or more accelerators to serve it. This is cluster territory.
That does not put it outside the scope of local AI. When we say local, we mean self-hosted, running on hardware you control, and that spans a range: from a model on a consumer GPU to a large open model on an organisation’s own cluster. The common thread is that you hold the weights, your data stays on your infrastructure, and you are not dependent on someone else’s API. Kimi K3 is the server-class end of that range, and for an organisation with the hardware, having an open model that competes with the closed frontier is a meaningful option to have.
The context
Kimi K3 is part of a clear pattern through 2026: strong open-weight models, several of them from Chinese labs, closing the gap with the closed frontier on reasoning and coding. It follows Moonshot’s Kimi K2 line and sits alongside releases like DeepSeek V4. For most people running models at home, the practical takeaway is the same as with other very large models: the smaller, runnable open models are what you will actually use day to day, while releases like this one shape the direction of the field and serve those with the infrastructure to run them.
We have added Kimi K3 to the model catalogue, with its hardware requirements and the licence uncertainty noted plainly.