Bitcoin, Ether, and all other major cryptocurrencies fell on Friday, with many pointing to a Beijing startup that distributed a model that surpassed Anthropic’s best at writing code.
Moonshot AI released Kimi K3 on Thursday, and by Friday morning, AI and semiconductor stocks were selling off across Asia as traders called it “Kimi Moment”, an echo of the DeepSeek shock that wiped out about $600 billion in Nvidia’s market capitalization in one session.
Moonshot built a 2.8 trillion parameter model with a context window of 1 million tokens, about four times the size of previous versions. This is done with an expert mixed design. That is, only a small slice of itself is activated for a given task, in this case 16 out of 896 experts. This is how we keep the cost of running such a large model low.
According to Moonshot’s technical blog, architectural changes have resulted in approximately 2.5x scaling efficiency over the previous generation.
Arena’s front-end code leaderboard was topped by K3 with 1,679 to Anthropic’s Claude Fable 5 with 1,631 and OpenAI’s GPT-5.6 with 1,618, taking first place in six of seven categories.
Moonshot’s own previous model was in 18th place. That’s the 17th jump in one release. In a broader test of general knowledge tasks, K3 ranks behind rather than ahead of the top Claude and OpenAI configurations, so this is a win in a specific area rather than overall.
