DeepSeek is a Chinese company specializing in large language models, systems capable of generating text, writing code or solving certain problems based on instructions. Based in Hangzhou, it serves the general public through its conversational assistant, as well as developers and businesses through its API and downloadable models. Its website, deepseek.com, provides access to its services and documentation.
Origins rooted in quantitative research
DeepSeek was founded in 2023 by Liang Wenfeng, also a co-founder of High-Flyer, a Chinese fund specializing in quantitative investing. The project benefits from this environment, where high-performance computing and machine learning play an important role. However, it differs from the fund’s financial activities in its objective: to develop general-purpose artificial intelligence models.
The company gained visibility with DeepSeek-V2 in 2024, followed by DeepSeek-V3 at the end of the same year. In January 2025, the release of DeepSeek-R1, focused on reasoning tasks, broadened its international audience. Its results then fueled debate over the resources actually needed to build high-performing models.
Open models, services and a drive for efficiency
DeepSeek’s activities span model design, training and release. Its systems can be used for writing, programming, document analysis and solving mathematical problems. The API allows these capabilities to be integrated into third-party applications, while the published weights make it possible to deploy certain models on infrastructure chosen by the user, subject to the necessary hardware capabilities and applicable licenses.
On the technical side, DeepSeek-V3 uses what is known as a “mixture of experts” architecture: only a subset of the parameters is activated for each processing task. This approach aims to limit the computation required. DeepSeek-R1 notably explores reinforcement learning to improve reasoning. The company publishes technical reports that allow its methods and evaluations to be examined.
This openness requires qualification: making weights accessible does not mean publishing all the data or the entire training process. Likewise, the training costs reported for a model do not necessarily represent all research, infrastructure and operating expenses.
What next?
DeepSeek’s trajectory will depend on its ability to maintain the quality of its models while ensuring reliable services. For businesses using them, selection criteria go beyond test results: data privacy, security, hosting arrangements, availability and regulatory compliance also weigh on decisions.
Its development is also taking place against a backdrop of US restrictions on exports of certain advanced chips to China. Software efficiency is therefore a strategic priority. For DeepSeek, the challenge will be to turn the visibility gained through its publications into lasting adoption, in a market where open models and proprietary offerings are evolving rapidly.