DeepSeek (深度求索, formally Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co.) is a Chinese AI company founded in Hangzhou in July 2023 and incubated by the quantitative trading firm High-Flyer. Its founder is Liang Wenfeng[1]. DeepSeek builds large language and reasoning models, and made its name with a combination that unsettled the industry: open model weights plus exceptionally low training costs. Its R1 release in January 2025 knocked global technology stocks down in a single session[3][4].
Unlike most frontier labs, DeepSeek ran on internal money from its parent rather than venture funding, until reports in April 2026 of a first external round at a valuation of at least US$10 billion[1]. It is also one of the few top-tier labs that publishes weights, technical reports and training details together[2][5].
🏢 July 2023: an AI company grown out of a quant fund
[ edit ]DeepSeek did not start in a garage. High-Flyer was founded in 2015, shipped its first AI model for quant strategies in 2016, passed RMB 10 billion under management in 2019 and RMB 100 billion in 2021[1]. Its "Fire-Flyer I" AI supercomputer went live in 2020, and in 2021 the firm committed RMB 1 billion to Fire-Flyer II — computing power that later became the foundation for training large models[1].
In July 2023 High-Flyer announced DeepSeek as a large-model company, moving into general artificial intelligence. Funding research with trading profits is what allowed it to produce frontier models for two years without raising outside capital[1].
👤 Liang Wenfeng: Zhanjiang to Zhejiang University, quant to AI
[ edit ]Liang was born in 1985 in Wuchuan, Zhanjiang, Guangdong, and entered Zhejiang University at 17, studying information and electronic engineering under Xiang Zhiyu with a focus on machine vision[1][7]. In 2008, while still a master's student, he was already leading a team applying machine learning to fully automated quant trading; he founded High-Flyer in 2015[1].
He rarely gives interviews but has repeatedly stressed "technical independence", and turned down investment from several major institutions and technology giants before DeepSeek became famous[1]. In March 2025 he entered the Hurun Global Rich List for the first time with RMB 33 billion; in April he met Nvidia chief executive Jensen Huang to discuss chip design[1].
🧪 V2 and V3: costs pressed to the floor
[ edit ]DeepSeek released V2 in May 2024, and its cheap API pricing set off a price war among Chinese model providers[1]. V3 followed on 27 December 2024, with its technical report posted to arXiv the same day: a mixture-of-experts model with 671B total parameters and 37B activated per token, using Multi-head Latent Attention and DeepSeekMoE, pre-trained on 14.8 trillion tokens[2].
The striking number was not a benchmark score but the bill: pre-training took 2.788 million H800 GPU hours, about US$5.576 million at the time[2]. The figure has been quoted ever since, because comparable American models were being trained on budgets measured in billions[3][7].
🌍 20 January 2025: R1 and the sell-off
[ edit ]On 20 January 2025 DeepSeek released R1, a reasoning model, and published the weights so others could build on it[1][4]. The same day, Liang spoke at a symposium of experts and business leaders chaired by Premier Li Qiang[1].
Within a week the mobile app went global: on 27 January it was the number one free app in the US App Store and in 51 other countries, with downloads jumping from one million to 2.6 million[4].
That trading day, the market re-priced the economics of AI. The Nasdaq fell 3.1% after the open, the S&P 500 1.85%, Nvidia 16.3%, ASML 7% and Siemens Energy around 20%[3]. Nvidia reportedly lost about US$560 billion in market value — the largest one-day loss in US stock market history — and briefly surrendered its crown as the world's most valuable company[7]. At Davos, Microsoft chief executive Satya Nadella said the new model was "an excellent AI advancement" and that China's progress had to be taken "very, very seriously"[7].
🏛 2025: from technical event to public event
[ edit ]After R1, DeepSeek moved from a quiet research shop to a public topic in mainland China: Liang appeared on the Hurun list, met Jensen Huang, and the team's R1 paper made the cover of a scientific journal in September 2025[1]. Attention shifted from "how is it so cheap" to "can the method be copied", especially while US export controls on advanced chips remained in place[3][7].
🚀 2026: V4 and a first outside round
[ edit ]In January 2026 DeepSeek published a paper, "Conditional Memory via Scalable Lookup", with Liang as a co-author[1]. On 24 April it released a V4 preview with open weights: the flagship V4-Pro has 1.6 trillion total parameters with 49 billion active per token, while V4-Flash has 284 billion total and 13 billion active, built around a one-million-token context[5].
V4-Pro reached general availability on 13 August, adding adjustable reasoning effort (low, high, max), native support for the OpenAI Responses API, and new pricing split into peak and off-peak rates — off-peak 50% cheaper — effective 16 August[6].
The same month, multiple outlets reported DeepSeek's first external fundraising: a valuation of at least US$10 billion and a raise of no less than US$300 million, with investors describing the deal as impossible to get into[1]. For a company that long advertised its independence from outside capital, it marked a change of strategy[1].
🧩 Products and how open it is
[ edit ]For general users there is a web app and a mobile app with "Expert" and "Instant" modes; developers use the API platform, whose documentation publishes model specifications and prices[5][6][8]. For researchers, weights are uploaded to Hugging Face alongside technical reports, so others can download, fine-tune and inspect the training method[2][5].
The "open weights plus cheap API" combination is the clearest difference from most frontier labs: OpenAI and Google largely ship API-only flagship models, while DeepSeek gives both — at the cost of letting competitors stand on its results[4][5].
📊 Why it is so cheap: three technical choices
[ edit ]The V3 report attributes the saving to three things. First, architecture: a mixture-of-experts design keeps total parameters large while activating only a slice per token (37B of 671B), cutting inference cost, paired with Multi-head Latent Attention to compress the memory cost of attention — both validated in V2[2].
Second, training stability: V3 uses an auxiliary-loss-free load-balancing strategy, avoiding the usual MoE trade-off between balanced expert usage and model quality, and adds multi-token prediction as a training objective for better data efficiency[2].
Third, precision and engineering: FP8 mixed-precision training lowers memory and bandwidth demands further while pre-training on 14.8 trillion tokens[2]. Together they produce the 2.788 million GPU-hour figure — not simply "cheaper chips"[2].
❓ FAQ
[ edit ]Q: When was DeepSeek founded?
July 2023, incubated by High-Flyer, headquartered in Hangzhou[1].
Q: Did R1 really cost only US$5.6 million?
The US$5.576 million figure is the pre-training cost of V3 (2.788 million H800 GPU hours) reported in its technical paper; R1 was built on that base with reinforcement learning, and the two numbers are often conflated[2][3].
Q: Are the models open source?
Weights are published on Hugging Face, including V3, R1 and the 2026 V4 line[2][5].
Q: Why did R1 hit Nvidia's share price?
The market assumed frontier AI required vast amounts of high-end GPUs; DeepSeek showed a cheaper path, forcing a re-rating of chip demand[3][4].
Q: Has DeepSeek raised money or listed?
It long relied on High-Flyer's internal funds; in April 2026 reports emerged of a first external round at a valuation of at least US$10 billion, not confirmed by the company[1].
📚 References
[ edit ]- Sina Finance: DeepSeek's first funding round reportedly above RMB 68 billion valuation (19 Apr 2026)
- DeepSeek-AI: DeepSeek-V3 Technical Report (arXiv:2412.19437)
- Reuters: China's DeepSeek sparks AI market rout (27 Jan 2025)
- TechCrunch: DeepSeek displaces ChatGPT as the App Store's top app (27 Jan 2025)
- DeepSeek API Docs: DeepSeek V4 Preview Release (24 Apr 2026)
- DeepSeek API Docs: DeepSeek-V4-Pro GA Release (13 Aug 2026)
- HK01: how DeepSeek did it (28 Jan 2025)
- DeepSeek official website
🎥 Videos
4 related videos · click to play
Play


