AICONTROLLED/LAB

JOURNAL / 2026.07.28

Moonshot publishes the full Kimi K3 model weights

The 2.8-trillion-parameter multimodal model can now be downloaded and deployed outside Moonshot; its scale, license, and published evidence define what that openness actually means.

  • Models
  • Releases
  • Open weights
  • Infrastructure

On July 27, Moonshot AI delivered the most important part of its Kimi K3 announcement: it published the model’s full weights, together with its configuration, license, and technical report. K3 had been available through the company’s products and API since July 16, but an API lets people use a service; weights let them study, modify, and run the model without sending every inference through the provider. That is the material change.

The official repository and Hugging Face distribution describe a multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters per token. It selects 16 of 896 experts at each step. It accepts text and images, offers a context window of just over one million tokens, and applies quantization from supervised fine-tuning onward to reduce inference cost. Compared with Kimi K2, which had 1.04 trillion total and 32.6 billion active parameters, this is not an incremental increase in scale.

Moonshot acknowledges that K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall performance, although its reported results place it close to both and ahead on some coding and tool-use evaluations. That table should not become a universal verdict. K3 is tested with Kimi Code on several tasks, while other models use Claude Code, Codex, or different harnesses; some results come from the lab itself and others from external leaderboards. The weights now allow third parties to repeat tests with the same harness and budget. They do not make the published comparisons retrospectively independent.

Open to inspection, demanding to run

The Kimi K3 license grants broad rights to use, copy, modify, distribute, fine-tune, and deploy the weights, including in commercial products. It is not without conditions. A model-as-a-service business whose aggregate revenue exceeds $20 million over twelve months must reach a separate agreement with Moonshot before using it commercially. Products with more than 100 million monthly active users or more than $20 million in monthly revenue must prominently display the Kimi K3 name.

That precision matters: “open weights” describes meaningful technical access, not the absence of restrictions or automatic equivalence with every definition of open source. An organization can keep data on its own infrastructure, adapt the model, and inspect its components without depending on the official API’s continued availability. It must also assess a custom license and take responsibility for the entire operation.

The physical barrier is higher still. Although the mixture-of-experts design avoids activating all 2.8 trillion parameters at once, Moonshot recommends a supernode configuration with 64 or more accelerators. Quantizations and third-party services will broaden access, but downloading a model does not mean running it on a workstation. In practice, openness has several levels: many people can inspect the artifacts; fewer can experiment with them at full scale; and fewer still can serve them with strong performance and reliability.

Public capability needs public limits

The launch materials include two unusually useful operational warnings. K3 can become unstable if an application does not preserve its full reasoning history between turns. Its training for long tasks can also make it “excessively proactive”: when intent is ambiguous or a minor problem appears, the model may make unexpected decisions on the user’s behalf. Moonshot recommends imposing explicit behavioral boundaries in the system prompt or files such as AGENTS.md.

That is a relevant admission for a model built around coding and tool use, but it does not replace a comprehensive safety evaluation. The public sources reviewed here explain architecture, deployment, benchmarks, and several product limitations; they do not present a separate card systematically measuring misuse, cybersecurity, deception, or loss of control. Now that the weights cannot be withdrawn from systems that download them, that evidence should be as reproducible as the capability tests.

My reading is that K3 genuinely expands who can research and build on a model near the frontier, but it does not by itself democratize the compute required to operate one. Its most durable contribution may be turning claims that could previously be tested only through an API into artifacts that other teams can retain, instrument, and contradict. Realizing that possibility will require shared evaluations, efficient implementations, and a fuller account of risk. Opening the weights is a condition for that scrutiny, not the scrutiny itself.

Sources

← Back to journal