Moonshot AI open-sources Kimi K3, a 2.8-trillion-parameter model, for download

Moonshot AI released the full weights of Kimi K3, its 2.8-trillion-parameter model, for public download on Hugging Face on July 27, 2026 – a week after the model debuted on benchmarks. Until now only its scores were public; now anyone can download and run it.

Kimi K3 is a Mixture-of-Experts model, an architecture that routes each token to a small subset of specialized sub-networks. It carries 2.8 trillion total parameters but activates only about 104 billion per token across 896 experts, with a 1-million-token context window and native text, image and video input. The weights use MXFP4 quantization and occupy roughly 1.4 terabytes.

Moonshot calls it the first open model to reach 2.8 trillion parameters, the world’s first open 3-trillion-parameter-class model, released under a custom license (the Kimi K3 License). Founder Yang Zhilin has said the company is chasing broad user growth through openness, a wager against the closed, paid models of US rivals such as OpenAI and Anthropic.

The strategy is already rattling China’s AI market. The benchmark debut a week earlier reportedly sent competitors’ shares lower, with Z.ai off as much as 30% in Hong Kong trading and MiniMax down as much as 16%.

Open weights are not the same as reproducible. Moonshot has not released training data or full training details, and the benchmark figures are the company’s own, not independently verified. At 1.4 terabytes, the model is also far too large for most organizations to run without serious hardware.

Whether developers build on Kimi K3 in numbers, or the download proves more statement than tool, is what to watch next.