Kimi releases K3 weights with 2.78T parameters, 1M context, and a new commercial license

AGI Hunt · wechat · 2026-07-28

Kimi has released the weights for K3, positioning it as the largest open-weight model so far.

K3 uses 2.78T total parameters with 104.2B activated, 16 of 896 experts active, a 1M-token context window, and native vision.
It ships under a new KimiK3License, not MIT. Commercial MaaS usage requires a separate agreement, and apps over $20M monthly revenue or 100M users must visibly show “KimiK3”.
Official and third-party API pricing is aligned at $3 per 1M input tokens and $15 per 1M output tokens, which appears designed to prevent price undercutting.
The MXFP4 weights are 1.4–1.5TB, too large for 8×A100 or 8×H200 nodes; 8×B300 is the first single-node setup that can fit them.
Benchmarks and estimates suggest K3 is cheaper per token on tasks like BrowseComp than GPT-5.6Sol and far cheaper than Fable5.
The report does not disclose training tokens or FLOPs, but estimates put training above 2e25 FLOPs and possibly 3–4e25, making it likely the largest model trained in China.
The model card notes weaknesses: reasoning traces may be cut off by agent frameworks, it may act too quickly in ambiguous situations, and its chat quality still lags Fable5 and Sol despite close benchmark scores.
The piece also frames K3 as a possible “fabless” model business model: free for self-use, but cloud service commercialization requires revenue sharing.

More from Infra