DeepSeek

What it is

DeepSeek (深度求索) is a Chinese AI research laboratory headquartered in Hangzhou, Zhejiang, China. Liang Wenfeng, co-founder of High-Flyer Quant, a Chinese quantitative trading firm, spun DeepSeek off from High-Flyer in July 2023 and serves as CEO of both companies. High-Flyer remains the principal backer. DeepSeek develops large language models spanning dense and mixture-of-experts (MoE) architectures, releasing most weights publicly under permissive licenses. As of July 2026, the lab runs 101 model repositories on Hugging Face. Its stated mission is fundamental AGI research, and it has operated for most of its history without outside investors.

History

DeepSeek shipped DeepSeek Coder on 2 November 2023, and the base LLM series on 29 November 2023. V2 (May 2024, 236 billion total parameters) introduced multi-head latent attention and MoE. V3 (December 2024) scaled to 671 billion total parameters with roughly 37 billion activated per token, trained on approximately 14 trillion tokens at a claimed cost of US$6 million, far below the US$100 million or more that US labs spent on comparable models.

The inflection point came on 20 January 2025. DeepSeek released R1, a reasoning model trained via reinforcement learning using Group Relative Policy Optimization on a V3 backbone. R1 matched or exceeded OpenAI o1 on several public benchmarks while being available as an open-weight download. Markets reacted sharply: on 27 January 2025, Nvidia's stock fell approximately 17%, erasing roughly US$589 billion in market capitalisation, the largest single-day market-cap loss in US stock market history at the time. The Nasdaq fell more than 3% that day. Shortly after R1's release, Liang Wenfeng participated in a symposium with Chinese Premier Li Qiang, signalling Chinese government support.

Current state

DeepSeek's frontier product as of July 2026 is the V4 family, previewed on 24 April 2026. V4-Pro carries 1.6 trillion parameters (49 billion active) and V4-Flash runs 284 billion (13 billion active); both carry 1 million-token context windows and are released as open weights on Hugging Face. V4-Flash uses hybrid attention that cuts FLOPs to roughly 10% of V3's at long context. The lab is raising its first outside capital: as of May 2026, it was in talks for a round at a US$45 billion valuation, led by China's state-backed China Integrated Circuit Industry Investment Fund, with cloud giants Tencent and Alibaba also in discussions. Liang Wenfeng holds approximately 84% ownership; he sought outside capital primarily to offer employees equity as competitor poaching intensified. The lab has adapted its training stack to work with Huawei Ascend chips, insulating it partly from US chip-export restrictions.

Relationships

DeepSeek is the most downloaded Chinese open-weight AI lab globally and the primary competitive reference for Chinese peers, including Zhipu AI's GLM-5.2 and MiniMax M3. Internationally, its open-weight strategy intensifies pressure on Meta AI's Llama series, while its low-cost training claims challenge the capital-intensive narratives of US closed-source labs. DeepSeek's efficiency methods, particularly MoE routing, RL-based reasoning, and low-cost training, have influenced both US and Chinese competitor designs. The open-vs-closed frontier debate is substantially shaped by DeepSeek's continued decision to release weights, which spreads Chinese AI research globally even as US export controls constrain its hardware options.

What to watch

Whether the US$45 billion funding round closes and how state-backed capital changes DeepSeek's research autonomy. Whether the V4 final release delivers on preview benchmarks against GPT-5.x and Gemini 3.x. How China's domestic chip ecosystem, primarily Huawei Ascend and domestic DRAM, constrains or enables the next model generation. Whether US regulators move to target open-weight Chinese AI releases specifically. How successfully Liang Wenfeng's equity programme retains researchers against offers from larger Chinese tech firms.