(Yicai) July 28 -- Moonshot AI has released the full weights for its new Kimi K3 artificial intelligence model, saying its 2.8 trillion parameters make it the world's first open-source model in the 3- trillion-parameter class.

The Beijing-based AI startup published Kimi K3's model weights on Hugging Face yesterday, alongside a detailed technical report and three core training infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. Released under a modified MIT licence, Kimi K3 can be downloaded, modified, fine-tuned, and deployed locally by developers.

In addition, Kimi K3's testing environment covered the Nvidia H200 and general-purpose graphics processing units from other vendors, suggesting that its deployment is not limited to the ecosystem of the US chip giant.

Kimi K3 has 104 billion active parameters and is based on the Mixture of Experts architecture, where each token activates only 16 out of a pool of 896 routing experts. It employs Moonshot AI’s proprietary Kimi Delta Attention hybrid linear attention mechanism along with Attention Residuals technologies, which natively supports visual understanding and has a context window of up to one million tokens. 

Compared with the previous generation, Kimi K2, the overall scaling efficiency has improved by about 2.5 times.

Kimi K3 outperformed all open-source and most closed-source models across dozens of benchmark tests in programming, agentic workflows, reasoning, vision, and other areas, according to the technical report. In Arena Intelligence's WebDev Arena, it achieved an Elo score of 1,679, becoming the first open-source model to claim the top spot on the leaderboard. 

In addition, Kimi K3 scored 57.1 on the Artificial Analysis Intelligence Index v4.1, ranking fourth worldwide. However, its overall performance still lags a little behind Claude Fable 5 and GPT-5.6 Sol, the report said.

The open source initiative aims to accelerate the deployment and popularization of cutting-edge AI, as well as to promote research in artificial general intelligence, Moonshot AI noted. With the simultaneous release of Kimi K3's complete weights, technical report, and core infrastructure technologies, open-source large language models have officially entered the 2.8-trillion-parameter stage, it stressed.

Editor: Martin Kadiev