company profile
DeepSeek: The Chinese Lab That Shocked AI
DeepSeek is a Chinese AI lab spun out of hedge fund High-Flyer. Its R1 model in January 2025 shocked the industry by matching frontier reasoning at a fraction of the training cost. Mostly API and free-tier today.
6 min read · Updated Apr 17, 2026
The lab nobody saw coming
For most of 2024, DeepSeek was a curiosity — a side project of a Chinese hedge fund, publishing open-weight models that were pretty good but not frontier. On January 20, 2025, it released R1 and the story changed overnight. R1 matched OpenAI’s o1 on multiple reasoning benchmarks, had open weights, and was reported to have been trained for under $10M. Markets convulsed. The lab has been unavoidable ever since.
Timeline
- 2016 — High-Flyer, the parent hedge fund, founded by Liang Wenfeng in Hangzhou.
- 2023 — DeepSeek formally spun out; early DeepSeek Coder and DeepSeek LLM releases.
- 2024 — DeepSeek V2 and V3; the efficient Mixture-of-Experts training recipe crystallizes.
- December 2024 — DeepSeek V3 ships; costs fraction of GPT-4 to train.
- January 2025 — DeepSeek R1 releases. US tech stocks drop $1T in one day.
- 2025 — Multiple follow-ups and distilled variants; widely adopted in open-source tooling.
- 2026 — DeepSeek continues to ship open-weight frontier models at low API prices.
Leadership
- Liang Wenfeng — Founder and CEO. Hedge fund background; funds the entire lab from High-Flyer’s balance sheet.
- The research team is deliberately small and young; DeepSeek has been public about hiring researchers rather than engineers.
Flagship products
- chat.deepseek.com — free consumer chat.
- DeepSeek V3, V3.1 — open-weight general models.
- DeepSeek R1 — open-weight reasoning model.
- DeepSeek-Coder — open-weight coding specialist.
- DeepSeek API — hosted inference at cut-rate prices.
Business model
DeepSeek is the outlier. It monetizes primarily through:
- API usage — very cheap, very high volume.
- Parent company subsidy — High-Flyer funds compute from trading profits.
There is no meaningful consumer subscription. The chat is free. The business is a strategic bet by a hedge fund on AI as a national and economic priority.
What makes DeepSeek distinct
- Training efficiency — V3 and R1 demonstrated that frontier models can be trained with radically less capital than Big AI assumed.
- Open weights — R1 is reproducible; anyone can download and self-host.
- Pricing pressure — DeepSeek’s API prices drag down the whole market.
- Geopolitical footprint — a Chinese lab delivering frontier results reshapes the AI-and-policy conversation.
Why this matters for picking a subscription
For end users, DeepSeek is mostly relevant in two ways:
- Free tier — chat.deepseek.com is one of the strongest free chatbots on the internet, especially for math, coding, and reasoning.
- Indirect effect — DeepSeek’s pricing pressure is why OpenAI, Anthropic, and Google have had to keep their Plus/Pro tiers competitive.
Enterprise and regulated users largely self-host the open weights rather than pay for DeepSeek’s hosted API.
Going deeper
Frequently asked questions
Who is behind DeepSeek?
DeepSeek is backed by High-Flyer, a quantitative hedge fund based in Hangzhou, China. The lab was formalized in 2023 and is led by Liang Wenfeng, High-Flyer's founder. It operates more like a research shop than a product company.
What happened in January 2025?
DeepSeek released R1, an open-weight reasoning model that matched or exceeded OpenAI's o1 on many benchmarks. The claim that it had been trained for a fraction of the compute cost triggered a one-day $1 trillion drop in US tech stocks.
Is DeepSeek free?
DeepSeek's consumer chat at chat.deepseek.com is free, and its API is priced far below OpenAI, Anthropic, or Google — often 10-30x cheaper per million tokens.
Is DeepSeek safe to use?
The models run on servers in China. For personal use that is usually fine; for regulated industries, Fortune 500 deployments, or anything sensitive, most US and EU customers prefer self-hosting the open-weight version rather than sending data through DeepSeek's hosted API.
What models does DeepSeek ship?
DeepSeek V3 and V3.1 (general-purpose), DeepSeek R1 and R1 derivatives (reasoning), DeepSeek-Coder (coding). All have open weights published on HuggingFace.
Can I run DeepSeek locally?
Yes — the full weights are on HuggingFace. R1 is very large (tens of billions of parameters) and needs serious hardware, but distilled versions run on a single high-end consumer GPU.