DeepSeek

A Hangzhou, China-based AI company and a series of open-source large language models. It shocked the global AI industry in 2025 with low-cost, high-performance models.

DeepSeek

Overview

DeepSeek (딥시크) is an artificial intelligence research company founded in July 2023 in Hangzhou, Zhejiang Province, China, and also the name of the large language model (LLM) series released by that company. The founder is Liang Wenfeng (량원펑, 梁文锋), a native of Guangdong Province, who is also the founder of the quantitative hedge fund High-Flyer. DeepSeek is regarded as a representative case that shook the global AI industry in early 2025 by achieving performance approaching that of top commercial models in the open-source camp at a very low training cost.

Main Content

Founding Background and Parent Company

DeepSeek began as the AI research organization of High-Flyer Capital. Liang Wenfeng had been securing large-scale GPU clusters since around 2021, and based on the capital and infrastructure accumulated through quantitative trading, he spun DeepSeek off as an independent legal entity in 2023. Unlike the typical venture capital investment model, this is a unique structure in which research is conducted with its own funds and its own computing resources.

Major Model Lineage

  • DeepSeek LLM (November 2023): The first released model, with 67B parameters. It aimed for balanced Chinese and English performance.
  • DeepSeek-V2 (May 2024): Introduced a MoE (Mixture of Experts) structure and MLA (Multi-head Latent Attention) to greatly reduce inference costs. From this point on, its "cost-effectiveness" image took hold.
  • DeepSeek-V3 (December 2024): A MoE model with 671B total parameters and about 37B activated parameters per token. It disclosed that training required about 2.788 million GPU hours (based on H800), drawing attention for costing about one-tenth of the industry's conventional wisdom.
  • DeepSeek-R1 (January 2025): A reinforcement-learning-based reasoning model. It claimed OpenAI o1-level performance in mathematics, coding, and logic, becoming a hot topic immediately upon release. Since then, many derivative models such as R1-Zero and distilled versions have been released.

Technical Characteristics

DeepSeek's core technical differentiator lies in efficiency. With the MoE architecture, only some expert layers are activated during inference, and MLA reduces KV cache usage. It also applied FP8 mixed-precision training and multi-token prediction (MTP) to simultaneously cut training and inference costs. The R1 series is known to use the GRPO (Group Relative Policy Optimization) technique, which boosts reasoning ability purely through reinforcement learning without separate large-scale supervised learning data.

Open-Source Strategy and Ecosystem

DeepSeek broadly permitted commercial use, including releasing model weights under the MIT license. Accordingly, global cloud providers (AWS, Azure, NVIDIA NIM, etc.) and companies at home and abroad have loaded DeepSeek models onto their platforms, creating an environment in which even small companies can fine-tune on their own servers.

Latest Trends

In January 2025, the DeepSeek app rose to No. 1 among free apps on the U.S. Apple App Store, triggering the "DeepSeek shock," in which NVIDIA's stock price plunged about 17% in a single day. This event raised questions about the conventional wisdom that enormous capital is essential for developing frontier models. Since then, DeepSeek has successively released follow-up versions of V3 and R1 (V3-0324, R1-0528, etc.), improving performance. In 2025, regulatory authorities in various countries raised data security and privacy issues, and some public institutions and companies have shown movements to restrict the use of DeepSeek. At the same time, in the open-source camp, many derivative models referencing DeepSeek's architecture have appeared, accelerating competition in low-cost, high-efficiency models. In Korea as well, there is a growing number of cases in which telecommunications companies, portals, and startups integrate DeepSeek models into their own services or release fine-tuned models that enhance Korean-language performance.

Related Topics

  • [[Large Language Models]]
  • [[Open-Source AI]]
  • [[Reinforcement Learning]]
  • [[NVIDIA]]
  • [[AI Semiconductors]]
  • [[Chinese IT Industry]]