DeepSeek is a Chinese artificial intelligence company focused on large language models, officially registered as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. Founded in 2023 and incubated by the quantitative trading firm High-Flyer, it is known for high-performance open-weight models trained and served at unusually low cost.
1 Company Profile
1.1 Name and Positioning
The Chinese name "深度求索" literally means "seeking the depths". The English name combines "deep" and "seek", echoing the company's focus on deep learning research. The firm positions itself as a basic-technology research lab, working on the training, compression and deployment of large language models and reasoning models.
1.2 Organizational Background
DeepSeek grew out of the AI research team at High-Flyer, a Chinese quantitative hedge fund that had built its own GPU clusters. After spinning off, the company kept that approach, running its own infrastructure and training framework rather than renting capacity at scale. Its headcount has remained small relative to peers with comparable model portfolios.
2 History
2.1 Founding
The company was established in July 2023. Its first releases came in November of the same year: DeepSeek Coder for programming tasks and DeepSeek LLM as a general-purpose model, both published together with weights and technical reports.
2.2 Architectural Turn and Price War
DeepSeek-V2, released in May 2024, introduced Multi-head Latent Attention and a self-developed mixture-of-experts structure, sharply reducing inference cost. Its API pricing undercut comparable products by a wide margin and triggered a wave of price cuts across the Chinese model market.
2.3 Scaling Up
DeepSeek-V3 arrived in December 2024 with a mixture-of-experts design of roughly 671 billion total parameters and about 37 billion activated per token. The disclosed training budget and GPU hours were far below what comparable frontier models were assumed to require, prompting widespread discussion about training efficiency.
2.4 Reasoning Models
In January 2025 the company released DeepSeek-R1, a reasoning model trained primarily with reinforcement learning on verifiable tasks rather than on large volumes of human-annotated reasoning traces. Weights were released under a permissive licence, and the accompanying app climbed app-store charts internationally.
2.5 Later Iterations
Subsequent releases followed at a brisk pace through 2025, including V3.1 and V3.2-Exp, which experimented with hybrid reasoning modes, longer context windows and sparse attention. Research behind R1 was also submitted for peer-reviewed publication.
3 Products and Technology
3.1 Model Lines
- V series: general-purpose language models covering chat, writing, coding and tool use.
- R series: reasoning models aimed at mathematics, code and multi-step problems.
- Coder series: models specialised in code completion and program understanding.
- Multimodal series: the Janus family, covering image understanding and generation.
3.2 Key Techniques
3.2.1 Mixture of Experts
Each token is routed to only a small subset of expert subnetworks, keeping compute per token manageable even as the total parameter count grows.
3.2.2 Multi-head Latent Attention
Key-value caches are compressed into a low-rank latent space, greatly reducing memory use during long-context inference — a prerequisite for cheap deployment.
3.2.3 Reinforcement Learning for Reasoning
The R1 line showed that outcome-based rewards alone, applied to tasks with checkable answers, can induce long chains of thought without supervised reasoning data.
4 Open Weights and Impact
4.1 Licensing
Most model weights are published with permissive terms allowing commercial use and redistribution, with additional conditions only for very large-scale commercial deployment. This has made the models among the most downloaded and most frequently fine-tuned in the open ecosystem.
4.2 Industry Effects
Cheap training plus open weights challenged the assumption that frontier capability requires enormous capital expenditure, and accelerated the spread of reasoning models among smaller teams. Public-sector bodies, cloud providers and hardware vendors in several countries have announced integration or on-premise deployment.
4.3 Criticism
Rising visibility brought scrutiny over training data provenance, distillation from other models, the boundaries of open release and safety alignment. Some regulators and competitors have raised compliance questions, to which the company has responded with technical reports and model documentation.
5 Applications and Ecosystem
5.1 Official Services
The company operates a web and mobile assistant, an open platform API, and developer features such as function calling and context caching.
5.2 Third-party Ecosystem
Because the weights are open, the community has produced quantised variants, local deployment toolkits and fine-tuned derivatives. Aggregator platforms also resell access, while enterprises mostly use the models for private deployment and internal knowledge-base question answering.