DeepSeek is a Chinese artificial intelligence laboratory based in Hangzhou, founded in 2023 and funded by the quantitative hedge fund High-Flyer, whose co-founder Liang Wenfeng leads the lab. It publishes frontier-scale large language models with open weights and detailed technical reports, and its January 2025 release of the R1 reasoning model triggered a global repricing of assumptions about the cost of frontier AI.1

Stated mission and record

DeepSeek's English-language website describes the company as "dedicated to exploring the essence of AGI" and says it will "embrace all who are passionate about AI."4 Its official Hugging Face organization profile carries a two-line credo: "Unravel the mystery of AGI with curiosity. Answer the essential question with long-termism."5 The lab publishes no formal charter or safety framework; the fullest statement of its principles is a pair of founder interviews with the Chinese outlet 36Kr, widely circulated in English translation (reported, translated), in which Liang Wenfeng frames curiosity-driven basic research and open publication as deliberate strategy rather than a commercial calculation.6

As of July 2026 the record stands as follows: DeepSeek-R1 shipped with open weights under an MIT license, DeepSeek-V3 and its successors disclosed training economics and infrastructure in unusually complete technical reports, and the release cadence documented at DeepSeek (model family) has made the lab a reference point for the disclosure culture surveyed at Chinese AI labs. Points of tension are recorded at DeepSeek R1 release shock: Western lab leaders disputed the reading of its cost figures (final-run versus total program cost) and alleged distillation from proprietary model outputs, claims DeepSeek did not confirm. The openness also has limits: weights, papers, and methods are published, but the training datasets themselves are not; the technical reports document data pipelines and token counts rather than releasing corpora.

History

High-Flyer had accumulated GPU clusters for quantitative trading research before US export controls tightened, and spun its AI research into DeepSeek. The lab's early releases (DeepSeek LLM, DeepSeek-Coder) tracked the open-weight state of the art; DeepSeek-V2 (2024) introduced multi-head latent attention and its DeepSeekMoE mixture-of-experts design, cutting inference costs sharply. DeepSeek-V3 (December 2024) reported frontier-class quality from a training run whose disclosed compute budget, roughly $5.6 million in GPU-hours for the final run as stated in its technical report, was an order of magnitude below Western peers' assumed spending.2

DeepSeek-R1 (January 2025) matched OpenAI's o1-class reasoning results, with open weights and a paper documenting its reinforcement-learning recipe;3 the release coincided with a historic one-day fall in AI-linked equities and pushed every frontier lab, from Anthropic to Google DeepMind, to respond publicly on efficiency. As of July 2026, reporting places its V4-generation models among the strongest open-weight systems; details on this wiki are verified per page.1

Research and models

DeepSeek's papers are unusually complete for a frontier lab: architecture, data pipeline, RL recipes, and infrastructure are documented, making the Transformer-based DeepSeek line, post-trained with RLHF-family and verifiable-reward methods, a reference point for the open research community alongside Meta AI's Llama and Mistral AI's catalog. Its models are standard entries on MMLU and reasoning leaderboards.2

See also

References


  1. Contemporaneous coverage of the January 2025 R1 release and market reaction (Reuters, Financial Times, January 2025). ↩↩

  2. DeepSeek-AI, "DeepSeek-V3 Technical Report," arXiv:2412.19437, December 2024. ↩↩

  3. DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning," arXiv:2501.12948, January 2025. ↩

  4. DeepSeek, English-language homepage, official website, as of July 2026. ↩

  5. DeepSeek, organization profile self-description, official Hugging Face organization page, as of July 2026. ↩

  6. Liang Wenfeng, interviews with 36Kr sub-brand Waves (暗涌), 2023 and July 2024; widely translated English versions circulated by outlets including ChinaTalk (reported, translated). ↩