Kimi K3 at 2.8 trillion parameters
Moonshot AI opened the weights of Kimi K3: 2.8 trillion parameters, 104 billion active per token, a million tokens of context.
Why it matters
The largest model whose weights can be downloaded, by the company's word the first open 3T-class model, and it is not American.
Of 896 routed experts, sixteen work on each token, alongside two shared ones; the model takes images natively. It launched through the API on 16 July; the weights were promised by 27 July and went up on Hugging Face on the 27th. The licence is its own: an operator selling models as a service whose group's aggregate revenue passes 20 million dollars in 12 months must sign a separate agreement with Moonshot before commercial use. It is 1.75 times the previous largest open model, DeepSeek V4 Pro at 1.6 trillion. In 2024 Meta called Llama 3.1 at 405 billion the largest openly available model; K3 is almost seven times larger.