ZICQ
中 Log in / Sign up
Newsroom Agentic #Kimi #Open Source Model #MoE Architecture #Agentic Intelligence #Reinforcement Learning

Kimi Open-Sources Kimi K2: Agentic Mixture-of-Experts Model Released

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 6 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Kimi K2, the latest Mixture-of-Experts (MoE) model from Kimi, features 32 billion activated parameters and 1 trillion total parameters, achieving state-of-the-art performance in frontier knowledge, math, and coding. Beyond answering questions, Kimi K2 is meticulously optimized for agentic tasks, executing complex workflows and tool interactions seamlessly. The model supports a 256K context length and introduces the MuonClip optimizer for enhanced training stability. With its advanced tool-callin


Kimi K2: A Breakthrough in Agentic Intelligence

Kimi K2, the latest Mixture-of-Experts (MoE) model from Kimi, boasts the following key features:

  • Large-Scale Parameters and Efficient Performance: With 32 billion activated parameters and 1 trillion total parameters, Kimi K2 excels in frontier knowledge, math, and coding.
  • Agentic Optimization: Beyond answering questions, Kimi K2 is capable of executing complex tasks such as generating interactive websites, planning travel itineraries, and automating code development.
  • Tool-Calling Capabilities: The model can understand and utilize various tools like IPython, search engines, and calendar applications without requiring developers to write complex workflows.

Technical Highlights

  1. MuonClip Optimizer: To address training instability, Kimi introduced the MuonClip optimizer. This optimizer stabilizes training by directly rescaling the weight matrices of the query and key projections, resulting in zero training spikes.

  2. Large-Scale Agentic Data Synthesis: Kimi K2's training data is generated by simulating real-world tool-use scenarios, covering hundreds of domains and thousands of tools. This approach enhances data diversity and provides high-quality samples for reinforcement learning.

  3. General Reinforcement Learning System: Kimi K2 employs a self-judging mechanism that provides rubric-based feedback for non-verifiable tasks. This method, combined with policy optimization for verifiable tasks, makes the model more robust in handling complex tasks.

Developer Support

  • Open Platform: Kimi K2 is now open-sourced, allowing developers to access and use the model through the Kimi platform.
  • Tool-Calling API: The Kimi platform offers a tool-calling API, enabling developers to easily integrate Kimi K2 into their applications.
  • Recommended Inference Engines: Kimi K2 supports multiple inference engines, including vLLM, SGLang, KTransformers, and TensorRT-LLM, allowing developers to choose the most suitable engine for deployment.

Industry Impact and Future Outlook

The release of Kimi K2 marks a significant milestone in agentic intelligence. Its powerful tool-calling capabilities and efficient training methods provide a more open platform for developers, driving the development of agentic applications. In the future, Kimi plans to add more advanced features such as thinking and visual understanding to Kimi K2, further enhancing the model's intelligence.

Developer Recommendations

  • Explore the Tool-Calling API: Developers are encouraged to try the tool-calling API of Kimi K2 to build smarter and more adaptable applications.
  • Stay Tuned for Updates: Since the vision features of Kimi K2 are not yet supported, developers should stay tuned for Kimi's future updates to gain access to more features.

Conclusion

The release of Kimi K2 showcases the great potential of agentic intelligence, and its open-source nature provides developers with a broader application space. As technology continues to advance, Kimi K2 is expected to play an important role in more fields.


Source: Kimi Blog (2026-09-07)

— END —

Tags: #Kimi #Open Source Model #MoE Architecture #Agentic Intelligence #Reinforcement Learning

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…