DeepSeek: History, Open-Weight Models & Pricing Timeline

A hedge fund's side project that outgrew it

DeepSeek is a Hangzhou lab founded in July 2023 by Liang Wenfeng, who also runs High-Flyer, the quantitative hedge fund that owns the company and funds it. Liang started buying Nvidia GPUs for High-Flyer's trading models years before export controls made that hard, and DeepSeek grew out of that hardware and the team around it. [2][3][4]

The company answers to High-Flyer's balance sheet rather than a venture syndicate's exit timeline, and it has stayed small: roughly 160 people as of 2025, building frontier-scale models with a fraction of the headcount of its Western counterparts. [3][4]

Atlas interpretation: That ownership structure explains a lot of what follows. DeepSeek ships models the way a research group publishes papers, on its own schedule, without the run-rate pressure that shapes pricing and marketing at a venture-backed lab. [3][4]

Training efficiency became a geopolitical argument

DeepSeek's V3 release in December 2024, 671 billion parameters trained at a reported cost of $5.6 million, forced a public argument about how much a frontier-class model should cost to build. [5]

Three weeks later DeepSeek released R1, open reasoning-model weights that its own benchmarks put near OpenAI's o1. Nvidia lost close to $600 billion of market value in a single session on the theory that cheaper training meant fewer chips would be needed, an episode documented as the DeepSeek selloff; the theory did not hold up over the rest of the year. [6][7]

Atlas interpretation: The efficiency framing was always a little too clean. DeepSeek's reported figure describes the final training run, not the compute, staff time and false starts that got it there, and outside estimates of the lab's total spend run higher. What the episode actually showed was that a small, self-funded lab outside the usual compute race could reach the frontier and publish the weights, a different and more durable claim than a dollar figure. [6][7][5]

Open weights, on a cadence

DeepSeek kept publishing through 2025 and 2026. V3.2 and V3.2-Speciale in December 2025 claimed gold-medal results at the IMO, the CMO and the ICPC World Finals under an MIT license, and V4 in April 2026 shipped the lab's first new architecture since V3, again downloadable. In between, DeepSeek-OCR treated a vision encoder as a compression scheme for text, an odd little release that read as an argument about the price of a long context window as much as an OCR tool. [8][9][10]

The price war it started and ended

DeepSeek's API pricing was part of the pitch from the start. In May 2026 it made that explicit: a promotional 75 percent discount on V4-Pro became the permanent list price, pushing input tokens down toward 44 cents per million. Three months later it reversed course: when V4-Pro reached general availability in August 2026, API prices rose by 50 to more than 1,100 percent, with peak and off-peak billing replacing the flat rate. [11][12]

Atlas interpretation: Both moves say more about demand than doctrine. The cut needed to compete against the open-weights wave from Alibaba and Zhipu AI; the increase followed once DeepSeek needed a peak-hour price signal to manage its own load. A lab that built its reputation on cheap inference started rationing it, which is what tends to happen once the traffic gets big enough to matter. [11][12]