DeepSeek — the origin story

Every origin story needs a hook. DeepSeek's is: a quantitative hedge fund, bruised by a market rout, quietly kept buying GPUs — and then spent them on AGI instead of alpha.

Act I — the quant years

Liang Wenfeng starts trading with machine learning in 2008 while still at Zhejiang University. In 2015 he co-founds High-Flyer with Xu Jin. The fund is famous for a geeky, quirky culture — mathematics and AI over meetings. By late 2017, AI makes most of the trading decisions. Assets grow past $12 billion.

Act II — the rout and the GPUs

The Chinese market turns. High-Flyer apologizes to investors, assets shrink by more than a third. Behind the scenes, Liang is doing the opposite of retrenching: he acquires 10,000 Nvidia A100 GPUs before US export restrictions close that door — stockpiled, as it later turns out, for a research lab that doesn't exist yet.

Act III — "a second start-up"

On April 14, 2023, High-Flyer announces an AGI research lab "unrelated to the firm's financial business." Zhengzhe Lu, High-Flyer's CEO, describes it that spring as "a new team independent of investment, what is equivalent to a second start-up": "We want to do things with greater value, things that go beyond the investment industry."

On July 17, 2023, that second start-up becomes DeepSeek. Venture capitalists pass — no near-term exit, they reason. High-Flyer funds the whole thing. The accidental business model becomes the famous one: no commercial pressure, no rigid KPIs, and the freedom to break with established model architecture.

Act IV — the models, and the moment the world noticed

• May 2024 — DeepSeek-V2, at a fraction of frontier costs.
• December 2024 — DeepSeek-V3: open, cheap, and startlingly good.
• January 2025 — DeepSeek-R1, a reasoning model with reasoning disclosed, plus the chatbot named DeepSeek. On January 27, 2025, global tech shares sell off. The 18-month-old hedge-fund spin-off is suddenly the story of the industry: high performance, comparatively little compute, open weights.
• 2025–2026 — V3.1, V3.2, and the long-context push.
• April 24, 2026 — DeepSeek-V4 preview: 1M-token context, hybrid Compressed Sparse Attention, Muon optimizer, FP4-quantized experts — Pro (1.6T total) and Flash (284B total / 13B active), open weights under MIT.
• July 31, 2026 — V4-Flash-0731: the official Flash API in public beta, re-post-trained, Responses-API native, Codex-adapted — and beating V4-Pro-Preview on agent benchmarks (Terminal-Bench 2.1: 82.7, DeepSWE: 54.4, Toolathlon-Verified: 70.3). A smaller model outperforming its own flagship, by post-training alone.

That is the build in the name of this site: the day DeepSeek shipped the fast, free, open model that wrote the page you're reading.

The company history: DeepSeek — company history · The founder: DeepSeek — the founder · The model here: deepseek flash 07/31.