Moonshot AI has released Kimi K2, a large language model (LLM) that has immediately ranked among the world’s top ten on the LMSys text arena leaderboard, outperforming rival open-source model DeepSeek.
The model, made available for free under a modified MIT license, is a one-trillion parameter mixture-of-experts system with a 128,000-token context window. It features 384 experts, with 32 billion parameters activated per task. Designed for AI agents, Kimi K2 targets autonomous problem-solving, reasoning and tool use across complex business and research applications.
“Kimi K2 scored higher than DeepSeek,” Moonshot AI said, highlighting its position among the strongest open-source systems currently available.
The training process combined real-world and simulated environments, while a self-judging mechanism allowed the model to assess the quality of its own outputs. Developers also introduced a new MuonClip optimizer to stabilize training, enabling the model to be pre-trained on 15.5 trillion tokens.
Businesses looking to deploy Kimi K2 will require at least 1TB of storage and clusters of 16 Nvidia H20 or H200 GPUs. Distilled versions for smaller-scale use are expected, following a similar rollout approach to DeepSeek. The model is available for download on the Hugging Face platform.
Source: Moonshot AI, Moonshot AI on Github
