K3
A large language model released by Moonshot AI that topped the frontend code arena leaderboard. Bloomberg reported that it challenges the perception of the United States holding a six-to-twelve-month lead over China in frontier AI development.
Core Facts
Timeline (last 90 days)
Ember-1, developed by Fairworks Research through post-training on K3, achieved the same success rate as K3 in online A/B testing of coding traffic while using 71% fewer reasoning tokens and 39% fewer total tokens.
Kimi Code Desktop is powered by Moonshot AI's K3 model.
Kimi (Moonshot AI)'s K3 model has 2.8 trillion parameters, activates approximately 104 billion parameters per token, supports one-million-token context, and has built-in vision.
K3 achieved 76.8% on SWE-bench, the best among open-source models, and can run 300 auxiliary agents in parallel.
K3 is trained on Kimi's proprietary sandbox platform Agent ENV, which only allows copying, saving, and rolling back sandboxes, and does not allow agents to build environments inside it themselves.
According to data from the OpenRouter platform, Moonshot AI's K3 series models generate approximately 300 billion tokens per day.
Usage of the K3 model has seen a slight decline in recent months.
A user's workplace has self-built and deployed the K3 and GLM 5.3 models, and has prohibited the use of third-party APIs in sensitive data scenarios.
The K3 large model released by Kimi has a parameter scale of nearly 3 trillion (2.8 trillion).
Kimi K3 supports an ultra-long context window of 1 million tokens.
All Facts (6)
The K3 large model released by Kimi has a parameter scale of nearly 3 trillion (2.8 trillion).
65%VerifiedKimi K3 supports an ultra-long context window of 1 million tokens.
65%UnverifiedEmber-1, developed by Fairworks Research through post-training on K3, achieved the same success rate as K3 in online A/B testing of coding traffic while using 71% fewer reasoning tokens and 39% fewer total tokens.
50%UnverifiedKimi Code Desktop is powered by Moonshot AI's K3 model.
50%UnverifiedKimi (Moonshot AI)'s K3 model has 2.8 trillion parameters, activates approximately 104 billion parameters per token, supports one-million-token context, and has built-in vision.
50%UnverifiedK3 is trained on Kimi's proprietary sandbox platform Agent ENV, which only allows copying, saving, and rolling back sandboxes, and does not allow agents to build environments inside it themselves.
50%