Open-weight release
Beijing-based Moonshot AI published the full model weights for its flagship Kimi K3 this week, completing a rollout that began when the model debuted at the World Artificial Intelligence Conference in Shanghai on July 16. Coverage on July 28 described the move as the largest open-weight AI model so far, a 2.8-trillion-parameter Mixture-of-Experts system available for download, modification and commercial use under a custom licence, with supporting infrastructure and a 47-page technical report released alongside it. Indian Express dated the formal announcement to Monday, July 27, while Tech Times reported the weights had already appeared on Hugging Face on July 26.
Technical specifications and license terms
Kimi K3 supports a one-million-token context window, activates 104 billion parameters from a pool of 896 experts at runtime, and handles native multimodal reasoning, according to details in the technical report. New architectural techniques include Kimi Delta Attention, Attention Residuals and Stable LatentMoE, which Moonshot said delivers roughly 2.5 times more intelligence per unit of compute than its predecessor Kimi K2. The model is hosted on Hugging Face, though its MXFP4 weights are estimated at around 1.4 terabytes, placing self-hosting beyond ordinary PCs and requiring a multi-GPU server setup, VentureBeat and Digital Trends reported.
The Kimi K3 licence, modelled on MIT but distinctly non-commercial, grants broad rights to researchers, startups and many enterprises while adding obligations for larger players. Companies earning more than $20 million annually that operate what Moonshot calls a "Model as a Service" business offering third-party inference or fine-tuning must obtain a separate commercial licence, and services exceeding 100 million users or $20 million in monthly revenue face additional display requirements.
Hardware ecosystem
Moonshot has not disclosed the AI chips used to train Kimi K3, but the model is already being adapted to domestic Chinese hardware alternatives, reducing reliance on Nvidia. Alibaba Group Holding's cloud unit confirmed its supernode, powered by proprietary Zhenwu M890 chips, is compatible with Kimi K3, while a computing platform running Huawei Technologies' Ascend 910C has also completed adaptation, according to the South China Morning Post.
US allegations and Chinese response
The release landed in the middle of a trans-Pacific dispute over Kimi K3's origins. On July 22, White House Office of Science and Technology Policy Director Michael Kratsios posted on X that the US government had "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." Treasury Secretary Scott Bessent followed with a warning that "when PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table." As of July 28, the Bureau of Industry and Security was formally investigating but no formal action had been taken, Tech Times reported.
China's Ministry of Commerce issued its first official response on Monday, translated by Georgetown's Center for Security and Emerging Technology, dismissing the accusations as lacking "any actual evidence" and "no legal backing" and characterising Washington's position as "a classic case of hegemonic behavior in the AI sphere." The ministry pointed to the pace of Chinese model development as evidence of genuine independent research investment, and asserted that "many US AI companies have distilled from China's models for their own R&D and training." Moonshot's head of enterprise business, Huang Zhenxin, separately told state-run media that K3's performance gains came from "original changes to underlying model architecture."
Industry division over open-weight policy
The dispute has sharpened a split among major US technology firms over how freely advanced AI models should be released. A letter published on July 24, urging Washington not to impose sweeping restrictions on downloadable models, reportedly doubled its signatories from 25 to 50 within a day, adding names including OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block and Ollama, with Nvidia CEO Jensen Huang backing the push. Amazon and Anthropic have not signed the letter, reflecting continued disagreement over whether open-weight AI should be constrained. With weights now available, the developer community can begin independent evaluation that was not possible during the model"s earlier API-only period, though fully replicating K3 will require substantial GPU memory.
Share this article
