Moonshot AI ships Kimi K3: 2.8 trillion parameters, the largest open-weight model ever released
In one sentence China-based Moonshot AI releases Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with native vision and a 1-million-token context window. It is the largest open-weight model published to date, with full weights landing July 26-27 under a Modified MIT license.
Picture an enormous library where, for every question, only the right shelf gets consulted: that is the Mixture-of-Experts principle, and Kimi K3 takes it to an unprecedented scale. Moonshot AI's model counts 2.8 trillion total parameters, but activates only about 104 billion of them per token — a giant that works selectively.
The real story is not just the size: the weights are open. Moonshot launched the model on July 16, then published the full weights days later (July 26-27) under a Modified MIT-style license. Anyone with suitable hardware can download it, study it, and run it on their own infrastructure without depending on anyone's servers.
For everyday users the effect is indirect but real: open models of this caliber push AI service prices down and let companies and researchers build alternatives to the big American labs. Early independent reviews place K3 at frontier level for agentic coding — autonomous, multi-step programming work.
There is a geopolitical angle too: Moonshot shows that, despite US chip export restrictions on China, Chinese labs can produce frontier-class models and hand them to the world for free — a strategy that puts pressure on labs selling closed models at premium prices.
Companies
Moonshot AI
Tools
Kimi K3, vLLM
Tags
Sources