China's Moonshot AI released Kimi K2, a sparse mixture-of-experts model with a trillion parameters in total and about 32 billion active per token, selecting 8 of 384 experts, with a 128,000-token context. The weights shipped under a modified MIT licence. What set it apart was the tilt toward agentic work — using tools and carrying multi-step tasks through on its own — trained by synthesizing large numbers of tool-use scenarios in post-training. The reaction repeated the shock of DeepSeek: near-frontier capability you could download rather than rent. An update in September strengthened coding and widened the context to 256,000 tokens.