Moonshot's Kimi K3 completes its open-weight release at 2.8 trillion parameters, the largest openly downloadable model to date
Kimi K3 arrived on 17 July at 2.8 trillion parameters and Moonshot moved to fully open-source it through late July, putting the largest open-weight model ever released into general hands — roughly 75% larger than DeepSeek V4-Pro at ~1.6 trillion.
The parameter count is the headline and the least interesting part. What a 2.8-trillion-parameter open release actually does is move the constraint from access to serving: anyone may download it, and almost nobody can run it. Openness at this scale is a statement about who can participate, and the answer is narrower than the licence implies.
The strategic read is that the open-weight frontier is now set in China and matched elsewhere. Llama, DeepSeek, Qwen and Kimi occupy the open-weight tier together, and the largest entry in it is Chinese. Whatever the policy debate concludes about open weights, the empirical position is already established.
For H2 2026 the practical consequence lands on inference providers rather than labs. A model this size is a hosting business before it is a research artifact, and the value accrues to whoever can serve it economically. That is the same rotation the funding market has been signalling all quarter.
Taskade — Open-Source LLM History: GPT-2 to Kimi K3 (2026) → · Layer3 Labs — Best Chinese AI Models 2026: Kimi K3, DeepSeek, Qwen →