Moonshot AI releases Kimi K2: A free, smarter rival to DeepSeek

Moonshot AI has released Kimi K2, an artificial intelligence large language model, for free under a modified MIT license. The LLM immediately ranked among the top ten most powerful AI models in the world on the LMSys text arena leaderboard. Kimi K2 scored higher than DeepSeek, another free AI that gained worldwide attention for its power and free licensing when it was released at the end of 2024.

Kimi K2 is a one-trillion parameter (1T) mixture-of-experts (MoE) model with a 128K context window and 384 experts using a 32-billion activated parameter subset. The AI was developed for use in AI agents specializing in autonomous problem-solving, reasoning, and tool use, allowing for its use in solving complex tasks and researching answers to high-level business problems.

Due to limited real-world tool-use training data, Kimi K2 was trained using a combination of real-world and simulated environments. Additionally, the AI training used a self-judging mechanism, allowing it to determine on its own whether its completed tasks were done adequately or not during training. The MuonClip optimizer was developed to address the training stability issue found in the Muon optimizer for neural networks, allowing Kimi K2 to be successfully pre-trained on 15.5T tokens quickly.

Readers interested in developing Kimi K2 for business use will need at least 1TB of storage for the model and a cluster of at least 16 Nvidia H20/H200 GPUs before downloading it freely from Hugging Face. Home users can easily run distilled versions of DeepSeek on Nvidia GPUs with 12GB or more of memory, like this card on Amazon, while waiting for distilled versions of Kimi K2.

Consumers can chat with the Kimi chatbot for free here, and software developers can pay to use Kimi through its API here.