In July 2025, the Chinese lab Moonshot AI released Kimi K2, an open-weight mixture-of-experts model with one trillion total parameters and 32 billion activated per token. The official model card describes it as “a state-of-the-art mixture-of-experts (MoE) language model” that was “trained with the Muon optimizer” and “achieves exceptional performance across frontier knowledge, reasoning, and coding tasks while being meticulously optimized for agentic capabilities.” It was released under a modified MIT license.
Unlike many large models tuned mainly for chat, Kimi K2 was “specifically designed for tool use, reasoning, and autonomous problem-solving,” with strong tool-calling built in. That positioned it as an open foundation for building agents that plan and act, rather than just answer questions.
Kimi K2 mattered because it pushed open-weight models to a trillion-parameter scale while keeping inference affordable through sparse activation, and because it underscored how quickly Chinese labs like Moonshot, DeepSeek, and Alibaba were releasing frontier-class systems openly. For businesses, it expanded the set of self-hostable models capable of serious agentic work.