DeepSeek releases R1, and the market notices

DeepSeek, a Chinese lab funded by the quantitative hedge fund High-Flyer, released R1: a reasoning model reported to be comparable to OpenAI’s o1 on mathematics, coding and reasoning benchmarks, published under an MIT licence with downloadable weights and an accompanying paper describing the method.

The method was the substance. The paper described R1-Zero, trained by reinforcement learning against verifiable answers with no supervised fine-tuning stage at all, and showed the model spontaneously developing longer deliberation and self-correction as training progressed. Where OpenAI had hidden its chains of thought and disclosed no technique, DeepSeek published both. It also released distilled versions small enough to run on a laptop.

The economic claim travelled further than the technical one. DeepSeek had stated that the preceding V3 base model cost about $5.6 million in its final training run — orders of magnitude below the budgets assumed for frontier models — and had achieved it while operating under US export controls that restricted its access to Nvidia’s best chips. The figure excluded research, staff and prior runs, and was widely misreported as the total cost of building R1, but the direction was not seriously disputed.

On 27 January, as the free DeepSeek app reached the top of the US App Store, US technology stocks sold off sharply; Nvidia fell about 17% in a single session, then the largest one-day market-capitalisation loss in history. The argument that frontier capability required enormous and defensible capital expenditure had been the premise of a great deal of investment, and R1 punctured it publicly.

It reset several assumptions at once: that Chinese labs trailed by a year or more, that export controls would hold the gap open, and that open weights would stay a step behind the closed frontier.