Moonshot AI Releases Kimi K3 Weights: the Largest Open-Weight Model Ever Published
The 2.8-trillion-parameter Mixture-of-Experts model lands on Hugging Face under a custom license — with a 1.4TB download, a $20 million revenue carve-out, and distillation accusations still swirling.
In Brief
- Moonshot AI published Kimi K3’s model weights on Hugging Face under a custom Kimi K3 License, alongside its technical report and key infrastructure code
- K3 is a 2.8-trillion-parameter Mixture-of-Experts model with a 1M-token context window whose weights occupy roughly 1.4 terabytes in MXFP4 format
- Moonshot claims the architecture delivers 2.5x more intelligence per unit of compute, though independent analysts caution against declaring parity with US frontier models
Chinese AI lab Moonshot AI released the model weights and technical report for Kimi K3 on Monday, publishing the weights on Hugging Face under a custom “Kimi K3 License” and open-sourcing parts of the infrastructure that powers it — including high-performance attention kernels, a Mixture-of-Experts communication library, and tools for running AI agents at scale, according to The Decoder.
“Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window,” Moonshot wrote in its announcement. “New model architecture: 2.5x the intelligence per unit of compute, not just more params.”
The release converts K3 from an API-only sensation into the largest open-weight model ever published — and lands eleven days after the model’s debut shook the frontier race and stretched Moonshot’s own GPU capacity past its limits within 48 hours.
Kimi K3 open weights come with a 1.4TB catch
Opening the weights does not mean most people can run them. The New Stack reports the download occupies roughly 1.4 terabytes in MXFP4 format, and realistic self-hosting requires a distributed environment of eight or more servers, each packing eight Nvidia H100 or B200-class accelerators. Moonshot’s documentation pitches the model at “long-horizon coding and end-to-end knowledge work,” served through an OpenAI-compatible API for everyone else.
The license is its own story. Community analyses relayed by Techmeme describe an MIT-inspired text with a commercial carve-out: companies earning over $20 million a year need a separate agreement with Moonshot. Researcher Nathan Lambert called it “inspired by MIT but distinctly non-commercial” for large players, while Peter Gostev summarized it as “MIT + these 2 points,” underscoring that the mechanism — a revenue threshold triggering a commercial deal — matters more than the label.
Performance claims remain contested. The New Stack’s hands-on comparison suggests K3 can match Anthropic’s Claude Fable 5 on programming tasks at roughly a third of the cost, but runs about four times slower. An independent UK Cyber Institute test found its cyber capabilities lag far behind frontier models, and its math skills trail similarly.
Distillation accusations shadow the release
The open-weight drop arrives with Washington watching. The White House has accused Moonshot of distilling Anthropic’s model to build K3 — Anthropic’s Sarah Heck called it IP theft, and OSTP Director Michael Kratsios alleged reliance on Anthropic models. Moonshot denies the charge, attributing K3’s gains to architectural innovations it calls Kimi Delta Attention and Attention Residuals.
Analysts have found reasons for skepticism on both sides. “While K3 constitutes progress, we’d hesitate to ascribe it near-parity with American frontier models, such as Fable 5, in actual tasks,” Morningstar senior equity analyst Malik Ahmed Khan wrote in a research note quoted by The New Stack.
Either way, the release raises the stakes in the open-weight fight now dividing the industry, days after Microsoft, Nvidia and Meta publicly defended open models while OpenAI and Anthropic pushed for restrictions. By publishing not just weights but the serving infrastructure around them, Moonshot has made K3 the reference point that debate will now orbit.
FAQ
What did Moonshot AI actually release?
The full model weights for Kimi K3 on Hugging Face, its technical report on GitHub, and open-source infrastructure including attention kernels, an MoE communication library, and agent-serving tools.
Can individuals run Kimi K3 locally?
Realistically no. The weights occupy about 1.4TB in MXFP4 format and self-hosting requires eight or more servers with eight H100 or B200-class GPUs each, per The New Stack.
What does the Kimi K3 License allow?
Community readings describe it as MIT-inspired with a carve-out requiring companies earning over $20 million a year to sign a separate commercial agreement with Moonshot.
