With Kimi-K2, Moonshot AI presents a new state-of-the-art language model specifically designed for agentic applications and demanding tool usage. The combination of huge MoE design, innovative MuonClip optimizer and large-scale RL training makes Kimi-K2-Instruct one of the most powerful open-source models for complex tasks.
Kimi-K2-Instruct
Moonshot AI
July 2025
Modified MIT license (commercial use permitted)
Mixture-of-Experts (MoE) language model
1 trillion (of which 32B active per token)
61 layers, 384 experts, 8 experts per token, MLA Attention, SwiGLU activation
160k vocabulary
128,000 tokens
We would be happy to advise you individually on which AI model suits your requirements. Arrange a no-obligation initial consultation with our AI experts and exploit the full potential of AI for your project!
Kimi-K2-Instruct was trained on an exceptionally broad database, including 15.5 trillion tokens as well as specially synthesized data for tool usage and agentic tasks. The pre-training was complemented by a comprehensive reinforcement learning process covering both verifiable and non-verifiable tasks.
With the specially developed MuonClip Optimizer, the scaling of this large-scale MoE model could be carried out stably and efficiently – a crucial basis for reliable use in real, complex scenarios.
Is Kimi-K2-Instruct the right AI model for your individual application? We will be happy to advise you comprehensively and personally.
Whether as an intelligent AI assistant, for automated code generation or for integration into scientific systems: Kimi-K2-Instruct provides the necessary architecture, performance and openness for productive use. Our team of experts will support you with selection, fine-tuning and hosting – fully managed in our German GPU Cloud if required.
