DeepSeekTranslation site

DeepSeek

DeepSeek is a Chinese artificial intelligence company based in Hangzhou that specialises in developing large language models (LLMs) and AI agents. Backed by the quantitative hedge fund High-Flyer, DeepSeek gained international attention for its open-weight models, especially the DeepSeek-R1, which achieved performance levels comparable to GPT-4 while significantly reducing costs. This success has made waves in both the AI and semiconductor industries.

Founded in 2023, DeepSeek is dedicated to researching cutting-edge general artificial intelligence models and technologies. The company tackles some of the most advanced AI challenges by leveraging its proprietary training framework, an intelligent computing cluster, and extensive computational resources. In just six months, the DeepSeek team has not only released but also open-sourced several large-scale models, each containing hundreds of billions of parameters. These include the DeepSeek-LLM, a general-purpose large language model, and the DeepSeek-Coder, which focuses on code generation.

Key facts

1. Full name: Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.
2. Founded: July 17, 2023
3. Headquarters: Hangzhou, Zhejiang, China
4. Founder / CEO: Liang Wenfeng (also co-founder of hedge fund High-Flyer)
5. Notable products: DeepSeek Coder, DeepSeek-V series, DeepSeek-R1 reasoning models, mobile chatbot app

Background and ownership

DeepSeek was spun out of the quantitative hedge fund High-Flyer, which had been using deep-learning-based trading systems for years. The fund’s co-founder, Liang Wenfeng, created DeepSeek as a dedicated LLM lab, funded and owned primarily by High-Flyer, with Liang holding a large personal stake via holding companies.

Technology and models

DeepSeek focuses on “open-weight” models: it publishes model parameters and technical reports, while often applying some use-case restrictions rather than fully permissive open-source licenses. Its model family includes DeepSeek Coder (code-focused LLMs), the DeepSeek-V/V2/V3 series, and DeepSeek-R1, a reasoning-optimised model trained with reinforcement learning. R1 and later variants are designed to emulate “thinking” traces and match or approach frontier models on benchmarks, while being relatively efficient to train and run.

Cost, disruption, and market impact

DeepSeek claimed that training its V3 model cost on the order of a few million US dollars and used significantly fewer GPUs than rivals. Even accounting for critiques that the true total cost was much higher, its efficiency shocked analysts and competitors. After the January 2025 launch of the DeepSeek chatbot app and R1 model, markets reacted sharply, including a major one-day drop in Nvidia’s valuation, as investors reassessed the economics of AI hardware demand.

Position in the global AI race and controversies

DeepSeek’s rapid rise is often framed as a “Sputnik moment” in the U.S.–China AI competition, demonstrating that a relatively small Chinese startup can produce models competitive with those from OpenAI, Meta, and Google. At the same time, it has drawn scrutiny: reports and a recent complaint by Anthropic allege that DeepSeek and other Chinese firms improperly harvested proprietary model interactions to improve their systems, which DeepSeek has not publicly confirmed in detail.

Chinese regulatory expectations and political sensitivities also shape the company’s deployments. Analyses of DeepSeek-R1 note that, while technically strong, it tends to align with official narratives on topics sensitive to Beijing, reflecting content controls common across major Chinese AI systems.