Moonshot AI has published Kimi K3's full model weights and a 47-page technical report. This closes the main gap left by the model's July 16 launch. K3 had been available through Moonshot's products and API, while the files needed to host it independently were still held back.
The public repository contains 96 weight files totaling 1.56 terabytes.
The model did not become smarter with the download. What changed is who can hold, test and serve it.
Moonshot's custom license permits use, copying, modification, distribution, fine-tuning and sale. It also has meaningful conditions. If a company or affiliate operates a model-as-a-service business and its aggregate revenue exceeds $20 million over any consecutive 12 months, it needs a separate Moonshot agreement before using K3 commercially. Commercial products with more than 100 million monthly users or $20 million in monthly revenue must display “Kimi K3” prominently. This is an open-weight release, not an unqualified open-source grant.
Portable does not mean personal
K3 is a sparse mixture-of-experts model with 2.8 trillion total parameters. It uses 16 of 896 routed experts, plus two shared experts, for each token. A host still needs the full checkpoint, including experts not used for that particular token.
The 1.56-terabyte size is consistent with more than a simple four-bit copy of every parameter. Moonshot's report says the routed expert weights use MXFP4, while attention layers, shared experts, routers and other parts remain at higher precision. The repository also includes the scales and metadata needed to interpret those values.
Moonshot recommends a “supernode”: a tightly connected cluster of 64 or more AI accelerators such as GPUs. That is company guidance, not a demonstrated hard minimum. The viable hardware configurations will depend on memory, networking and serving software.
That software was still arriving with the weights. When I checked minutes after release, the model card's links to vLLM and SGLang deployment recipes both returned 404 pages. The vLLM support change and SGLang support change were open and unmerged. This may be a short rollout lag. It still shows the practical boundary: files can be public before the advertised path for serving them is ready.
Most people will not run K3 on a laptop or a typical GPU server. Practical self-hosting is aimed at organizations with large, tightly connected accelerator clusters.
What the files make testable
The checkpoint lets outsiders inspect the released configuration and evaluate the same downloadable artifact under their own software. That should produce cleaner comparisons than Moonshot's launch table, which mixes different agent programs and score sources.
It does not validate the company's training or benchmark claims by itself. Independent hosts must still establish whether the checkpoint behaves like Moonshot's hosted service.
One deployment caveat matters immediately. K3 expects the software around it to pass the model's earlier reasoning back into later turns. Moonshot warns that quality can become unstable if the wrapper omits that history. Independent tests should therefore report the serving engine and conversation wrapper they use, not just the model name.
Hosts can now compare the downloadable checkpoint with Moonshot's API, including quality, long-context behavior and serving cost on their own hardware. The release makes those comparisons possible; it does not supply the results on release day.
What actually opened
Before public release, Moonshot and approved deployment partners controlled access to the checkpoint. Most users reached K3 through Moonshot's products or API. A public download can be copied and mirrored, so an API or pricing decision by Moonshot cannot by itself remove every copy. Licenses, laws and hardware access can still constrain who may run or distribute it.
The release moves control from one provider to a wider group of well-equipped operators. It does not put frontier-scale inference in everyone's hands.
Source graph: Semble source collection