The company Moonshot AI officially released its latest masterpiece - the Kimi K2 model, and simultaneously announced it as open source. This foundational model based on the MoE architecture has attracted widespread attention in the AI field since its release, thanks to its strong coding capabilities and excellent general Agent task processing capabilities.
The Kimi K2 model has a total of 1T parameters, with 32B activated parameters. It achieved top performance among open-source models in a series of benchmark tests such as SWE Bench Verified, Tau2, and AceBench, fully demonstrating its leading strengths in code writing, Agent task execution, and mathematical reasoning.
In the pre-training phase, Kimi K2 adopted an innovative MuonClip optimizer, which effectively solved the problem of large attention logits during large-scale training, significantly improving training stability and token usage efficiency. The Moonshot AI team successfully completed a stable training of 15.5T tokens, without any loss spike throughout the process, providing new insights for stable and efficient training of trillion-parameter models.
Beyond its excellent performance in benchmark tests, Kimi K2 also demonstrated strong capability generalization and practicality in real-world applications. In terms of coding ability, Kimi K2 not only generates front-end code that combines design sense and visual expression, supporting complex expressions such as particle systems, visualization, and 3D scenes, but can also automatically build a complete futures trading interface without specific instructions, showcasing its powerful autonomous programming ability.
In terms of Agent tool calls, Kimi K2 also performs excellently. It can stably parse complex instructions, automatically decompose requirements into a series of standardized, directly executable ToolCall structures, seamlessly integrate with various Agent/Coding frameworks, and complete complex tasks or automated coding. Whether analyzing the impact of remote work ratio on salary or developing a fan- chasing plan for Coldplay and completing related planning, Kimi K2 can easily cope with them, demonstrating its strong Agent capabilities.
Additionally, Kimi K2 has made significant improvements in stylistic writing. Whether rewriting scientific texts in a middle school student's tone or imitating Apple ad copy, Kimi K2 can accurately control the output style while preserving the original meaning and expressive style. In fictional writing tasks, the text generated by Kimi K2 pays more attention to details and emotions, no longer abstract and general, providing users with a richer creative experience.
Moonshot AI not only released the Kimi K2 model this time, but also simultaneously open-sourced two model versions: Kimi-K2-Base and Kimi-K2-Instruct. Among them, Kimi-K2-Base is a base pre-trained model that has not undergone instruction fine-tuning, suitable for research and custom scenarios; while Kimi-K2-Instruct is a general instruction fine-tuned version, performing outstandingly in most Q&A and Agent tasks. The model and fp8 weight files have been open-sourced on the HuggingFace platform, available for free use by developers.
To facilitate developers' deployment and use, inference engines such as vLLM, SGLang, and ktransformers have also synchronized support for the Kimi K2 model. Developers can deploy it on their own servers and obtain the same experience as the Kimi open platform API.
In terms of API services, Kimi K2 also provides comprehensive support. Its API service is now fully launched, supporting up to 128K context length, with stronger generality and tool call capabilities. The billing plans are flexible and reasonable, costing only 4 yuan per million input tokens and 16 yuan per million output tokens. It is also compatible with both OpenAI and Anthropic API formats, allowing developers to switch seamlessly.