ReleaseMoonshot AI
Moonshot AI Launches Kimi K2 Open-Source Model for Agentic Intelligence
Moonshot AI has released Kimi K2, a groundbreaking open-source large language model built on a Mixture-of-Experts (MoE) architecture with one trillion total parameters and 32 billion activated parameters. Trained on over 15.5 trillion tokens, Kimi K2 excels at complex task decomposition, native tool integration, coding, and frontier knowledge reasoning—outperforming leading open-source peers like DeepSeek’s V3 and rivaling proprietary U.S. models in key benchmarks. Two variants are available: Kimi-K2-Base for researchers and developers seeking full fine-tuning control, and Kimi-K2-Instruct for drop-in, general-purpose chat and agentic workflows. Released under a permissive MIT-style license, the model is freely accessible via web and mobile applications, inviting the global developer community to innovate, extend, and deploy advanced AI capabilities without barriers.
The idea behind it
Where it came from