DownloadKimi K2 (K2-Instruct-0905) is a state-of-the-art Mixture-of-Experts (MoE) language model developed by Moonshot AI, designed for high-performance reasoning, coding assistance, and agentic task orchestration. It features 1 trillion total parameters with 32 billion activated per token, enabling strong efficiency while maintaining very high capability. Kimi K2 demonstrates major gains in real-world coding and tool-use benchmarks, especially in SWE-Bench, Terminal-Bench, and multilingual programming tasks. Its 256K token context window allows it to handle extremely long workflows, multi-file repositories, conversations, and documents without losing coherence. The model is optimized for agentic intelligence, meaning it can autonomously decide when to call functions, perform actions, or use external tools during problem solving. Kimi K2 also provides enhanced frontend and UI coding support, generating practical, usable code rather than only algorithmic solutions.
Features
- 1T-parameter MoE architecture with 32B active parameters per token
- 256K context window for long-range reasoning and multi-file coding tasks
- High performance on agent tasks, including tool use and autonomous workflows
- Strong coding intelligence across frontend, backend, and full-stack development
- Optimized tool-calling capabilities for structured function execution
- Suitable for self-hosted inference using vLLM, SGLang, TensorRT-LLM, etc.
- SWE-Bench and Terminal-Bench leading performance in realistic software tasks
- Released under a Modified MIT License, enabling flexible research and deployment
Project Samples
