DeepSeek is a Chinese artificial intelligence lab, founded in 2023 and bankrolled by the quantitative hedge fund High-Flyer, that builds large language models it gives away for anyone to download, modify, and run. It matters because in January 2025 it showed the AI industry could get frontier-level results for a fraction of the money Silicon Valley was spending — a claim that wiped roughly $600 billion off Nvidia’s market value in a single day and forced every major AI lab to rethink how much compute a top model actually requires.
Who built it, and why
DeepSeek was founded in Hangzhou, China, by Liang Wenfeng, a quant trader who co-founded the hedge fund High-Flyer in 2016 and used AI to power its trading strategies. Around 2021, High-Flyer began stockpiling Nvidia GPUs for AI research, just ahead of US export restrictions that made those chips harder for Chinese firms to buy. In 2023, Liang spun the research effort out into a standalone company, DeepSeek, with an explicit goal of working toward artificial general intelligence rather than shipping a quick commercial product. Because DeepSeek was self-funded by an already-profitable hedge fund, it didn’t need outside investors or a subscription business for its first two years — a very different path from OpenAI or Anthropic, which raised billions before releasing a model.
The models: V3 and R1
DeepSeek’s breakthrough came from two releases. DeepSeek-V3, launched in December 2024, is a 671-billion-parameter model that only activates a fraction of those parameters for any given task — an architecture called mixture-of-experts, which keeps running costs down while preserving capability. DeepSeek said it trained V3’s final run for about $5.6 million, strikingly low next to the tens or hundreds of millions typically cited for comparable US models, by combining that architecture with efficient engineering tricks rather than brute-force computing power.
DeepSeek-R1, released in January 2025, took V3 and added large-scale reinforcement learning to make it “reason” through problems step by step before answering — the same idea behind OpenAI’s o1 models. R1 matched or beat those reasoning models on math and coding benchmarks, and DeepSeek released its full weights under the permissive MIT license, meaning anyone can download, fine-tune, or build a commercial product on top of it for free. That combination — near-frontier performance, radically lower reported cost, and a fully open license — is what made R1 different from a typical model release.
Why it rattled the industry
R1’s launch coincided with the app topping download charts on Apple’s App Store, and by January 27, 2025, investors had concluded that if a Chinese lab could reach near-frontier performance this cheaply, the enormous capital spending by Nvidia’s biggest customers might not be strictly necessary. That single-day repricing became the largest one-day market-cap loss for any company in US stock market history. The panic eased in the following weeks as details about R1’s actual training pipeline emerged and analysts noted the figures likely excluded earlier research costs, but the episode permanently changed how the industry talks about AI efficiency: it’s no longer assumed that only a handful of well-funded US labs can push the frontier forward.
DeepSeek’s models also demonstrated that export controls on advanced chips hadn’t stopped Chinese labs from competing — it pushed them toward more efficient use of the hardware they could still access, a pattern later echoed by other Chinese labs releasing competitive open-weight models.
What you can actually do with it
DeepSeek’s chat app and API are built around the same V3/R1 lineage, now continued in newer V4 models. Anyone can try the free web chat, and developers can call the models through an API that is compatible with the same request format used by OpenAI and Anthropic, documented on DeepSeek’s API quick-start guide. As of July 2026, per DeepSeek’s own pricing page, the budget-tier model runs about $0.14 per million input tokens and $0.28 per million output tokens, with a cheaper cached-input rate for repeated context — pricing that undercuts most Western frontier models by a wide margin, though it changes over time, so check the live page for current numbers.
That low price comes with a catch worth knowing before adopting it: DeepSeek’s servers and data handling are based in China, and its privacy policy discloses collection of account details, chat history, and device data. Citing those data-residency concerns, government agencies in Australia, Italy, South Korea, Taiwan, and elsewhere have restricted or banned DeepSeek on official devices — a consideration for any organization deciding whether to run it in production versus self-hosting the open-weight files instead.
In the news
DeepSeek is back in the headlines: it has opened talks for a new funding round that would value the company at $71 billion, roughly 40% above the valuation set by its first-ever outside investment round barely a month earlier. According to reporting cited in our brief on the talks, the new capital is meant to fund DeepSeek’s own data centers and chip purchases as it expands from chatbots into AI agents — a sign that even the lab famous for doing more with less now needs significantly more compute to keep pace.
FAQ
Is DeepSeek open source? Its model weights (V3, R1, and later releases) are published under the MIT license, so anyone can download and modify them. The training data and code used to produce those weights are not published, which is why they’re usually called “open-weight” rather than fully open-source.
Is DeepSeek free to use? The web chat is free. The API is metered per token, with prices well below most competing frontier models, though exact rates change — check the current pricing page before building on it.
Is it safe to use for sensitive data? Treat it like any cloud AI service based in a jurisdiction with different data-access rules than your own: fine for general use, but several governments have restricted it on official devices over data-residency concerns, and organizations handling sensitive data often prefer self-hosting the open weights instead of using DeepSeek’s own servers.
How is DeepSeek different from ChatGPT or Claude? All three are large language model chatbots, but DeepSeek’s models are open-weight and can be downloaded and run independently of the company, while ChatGPT (OpenAI) and Claude (Anthropic) are closed models accessible only through those companies’ own apps and APIs.