DeepSeek-R1

DeepSeek-R1 was released on 20 January 2025 with its weights, under an MIT licence, and a paper that explained how it had been made. Starting from the V3 base, the laboratory had applied reinforcement learning with rewards only for correct answers on maths and code, no human-written reasoning at all, and the model had learned by itself to think at length, to check its work and, in a passage the paper called an "aha moment", to stop and reconsider mid-solution. Its scores matched OpenAI's o1, whose method had been secret, and it cost a fraction as much to run.

The week that followed is the reason the event is at the top level of this timeline. DeepSeek's app reached the top of the American App Store; on Monday 27 January Nvidia's shares fell seventeen percent, erasing about $590 billion of value, the largest one-day loss any company had suffered, on the thought that if frontier models could be trained this cheaply the demand for chips had been overestimated. The thought did not last, and Nvidia was worth four trillion dollars by July, but the assumption that the frontier belonged to three American laboratories and their capital did not recover.

R1 also made the reasoning recipe public. Every laboratory's models thought out loud within months, and the distilled versions of R1 ran on laptops.