Kimi K3 is finally self-hostable: Moonshot releases 2.8T open weights

Moonshot AI has released the full Kimi K3 weights and technical report, turning last week's hosted-model launch into an open-weight release that developers can download, deploy, and inspect. The company describes K3 as the world's first open 3T-class model: a 2.8-trillion-parameter mixture-of-experts system with 104 billion active parameters, native text, image, and video understanding, and a 1-million-token context window. 1 2

That distinction matters. The July 16 release made K3 available through Moonshot's products and API; this release makes the model itself available under the Kimi K3 License. It is now a candidate for private deployments and custom serving, subject to the license and the hardware needed to run a model of this size. 3

K3 combines Kimi Delta Attention with Attention Residuals, then routes each token through 16 of 896 experts. Moonshot says those architectural and training changes deliver roughly 2.5× better scaling efficiency than Kimi K2. That is a company-reported efficiency claim, not an independent measurement. The model also uses MXFP4 weights and MXFP8 activations, while the release opens the infrastructure around the model rather than treating weights as a standalone artifact. 4

A same-day report identifies that infrastructure as FlashKDA, a high-performance Kimi Delta Attention kernel; MoonEP, a distributed MoE communication library; and AgentENV, a distributed environment system for large-scale agent workflows. The report attributes the release to the Kimi team and describes FlashKDA as a CUTLASS-based backend. 5

For hosted access, the model card points developers to the kimi-k3 model on Moonshot's API, with OpenAI- and Anthropic-compatible interfaces. For self-hosting, Moonshot recommends vLLM, SGLang, or TokenSpeed, and its technical blog recommends supernode deployments with at least 64 accelerators. In other words, the weights are public, but they are not lightweight. 2 4

Moonshot's model card reports 93.5 on GPQA Diamond, 84.8 on OSWorld-Verified, and 94.5 on MCPMark-Verified. The results use maximum reasoning effort and model-specific agent harnesses where noted, so they are useful launch evidence rather than a clean independent ranking. The technical blog itself says K3 still trails the strongest proprietary models in overall user experience. 2 4

The operational limitations are unusually important for developers. Thinking is always enabled, and multi-turn or tool-using applications must pass back the complete assistant message, including reasoning_content and tool_calls; dropping that history can make generation unstable. Moonshot also warns that K3 can be overly proactive on ambiguous tasks. The first serious tests will therefore be less about whether anyone can download the model and more about whether serving stacks can preserve its reasoning state at acceptable cost.

Read the model card · Read the technical report