DownloadKimi K2 Instruct is a high-performance Mixture-of-Experts (MoE) language model developed by Moonshot AI, activating 32B parameters per forward pass from a total 1 trillion. Designed for agentic reasoning, tool use, and advanced coding tasks, it achieves SOTA-level performance on multiple benchmarks such as SWE-Bench, AIME, and MMLU. Trained on 15.5T tokens using the Muon optimizer, it incorporates novel techniques for scaling stability. Kimi K2 supports a 128K context window, enabling detailed multi-turn conversations and long input handling. It includes native support for tool-calling, making it suitable for autonomous agents and real-world task execution. The Instruct variant is fine-tuned for chat-style interaction and general-purpose deployment, while the Base variant targets research and customization. Kimi K2 is released under a modified MIT license and deployable through engines like vLLM, SGLang, KTransformers, and TensorRT-LLM.
Features
- 1T parameter MoE with 32B active parameters per inference
- 128K context length for long-form tasks and reasoning
- Exceptional agentic performance and tool-calling capabilities
- Top-tier results on SWE-Bench, AIME, MMLU, and coding tasks
- Uses Muon optimizer for stable, scalable training
- Available in Instruct and Base variants
- Released under a modified MIT license
- Supports deployment via vLLM, SGLang, TensorRT-LLM, and more Preguntar a ChatGPT
Project Samples
