The Moon's Dark Side Corporation has taken a bold step forward in the realm of artificial intelligence with the official release and open-sourcing of their latest masterpiece, the Kimi K2 model. This monumental achievement is based on the Mixture of Experts (MoE) architecture, boasting an impressive 1 trillion parameters and 32 billion activation parameters. The Kimi K2 has already garnered significant attention within the AI community for its robust coding capabilities and exceptional performance in handling general Agent tasks.
In a series of benchmark performance tests, including SWE Bench Verified, Tau2, and AceBench, the Kimi K2 has consistently delivered top-tier results among open-source models, showcasing its prowess in code writing, Agent task execution, and mathematical reasoning. The model's pre-training phase was enhanced by the innovative MuonClip optimizer, which effectively addressed the issue of large attention logits in large-scale training, elevating training stability and token efficiency to new heights. The Moon's Dark Side team successfully completed a stable training of 15.5 trillion tokens without any loss spikes, offering a fresh perspective on the stable and efficient training of trillion-parameter models.
Beyond its benchmark performance, the Kimi K2 has demonstrated remarkable versatility and practicality in real-world applications. In terms of coding capabilities, the model is not only capable of generating front-end code that balances design sensibility with visual performance, supporting complex presentations such as particle systems, visualization, and 3D scenes, but it can also autonomously construct a complete futures trading interface without specific instructions, highlighting its robust self-programming capabilities.
The Kimi K2 also excels in Agent tool invocation, reliably parsing complex instructions and automatically decomposing them into a series of well-formatted, executable ToolCall structures. It seamlessly integrates with various Agent/Coding frameworks to complete complex tasks or automate coding. Whether analyzing the impact of remote work ratios on salaries or planning a fan-chasing itinerary for Coldplay enthusiasts, the Kimi K2 handles these tasks with ease, showcasing its formidable Agent capabilities.
In the realm of stylized writing, the Kimi K2 has also seen significant improvements. It can accurately control output style while retaining the original meaning and expression style, whether rewriting scientific texts in a tone suitable for middle school students or emulating Apple's advertising copy. In fictional writing tasks, the Kimi K2 generates text that pays more attention to details and emotions, offering users a richer creative experience.
The Moon's Dark Side Corporation has not only released the Kimi K2 model but also open-sourced two model versions: Kimi-K2-Base and Kimi-K2-Instruct. The Kimi-K2-Base is a basic pre-trained model without instruction fine-tuning, ideal for research and custom scenarios, while the Kimi-K2-Instruct is a general instruction fine-tuned version that excels in most question-answering and Agent tasks. Both models and their fp8 weight files are now available on the HuggingFace platform for developers to freely utilize.
For ease of deployment and use, inference engines such as vLLM, SGLang, and ktransformers have also been synchronized to support the Kimi K2 model. Developers can deploy on their own servers to achieve the same experience as the Kimi Open Platform API.
In terms of API services, the Kimi K2 provides comprehensive support. Its API service is now fully operational, supporting up to 128K context, offering enhanced generality and tool invocation capabilities. The pricing plan is flexible and reasonable, with only 4 dollars per million input tokens and 16 dollars per million output tokens. It also supports both OpenAI and Anthropic API formats, facilitating seamless switching for developers.