DeepSeek

DeepSeek is a Chinese artificial intelligence company headquartered in Hangzhou, China. It develops large language models and has drawn wide attention for releasing many of them as open-weight models, meaning their trained parameters can be downloaded and run by others. Its best-known release, the reasoning model R1, was published in January 2025 and was noted for achieving results comparable to leading proprietary systems at much lower reported training and usage costs. The company's stated long-term aim is artificial general intelligence.

History

DeepSeek grew out of High-Flyer, a quantitative investment firm co-founded in 2015 by Liang Wenfeng, a graduate of Zhejiang University. High-Flyer had used machine learning in its trading strategies and built large GPU clusters for that work. The AI research unit was announced in April 2023 and formally incorporated as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. in July 2023. Some secondary sources give May 2023 as the founding date, so the July date, which appears in the company's registration record and in most reference works, is the more widely accepted one.

Early releases included the DeepSeek-Coder and DeepSeek-LLM series in 2023. The mixture-of-experts model DeepSeek-V2 appeared in May 2024 and was notable for low API prices, which prompted a price war among Chinese AI providers. DeepSeek-V3 followed in December 2024 and was reported to rival GPT-4o on several benchmarks. R1 was released on 20 January 2025, and a consumer chatbot built on the model reached the top of Apple's App Store in the United States within days. The market reaction was large: reports at the time said Nvidia shares fell sharply in a single session, wiping out a very large sum of market value.

The V4 family was previewed on 24 April 2026. It comprised a large model, V4-Pro, with roughly 1.6 trillion total parameters and about 49 billion active per token, and a smaller V4-Flash. V4-Flash was released with open weights under the MIT license on 31 July 2026. V4-Pro reached general availability in mid-August 2026, with sources giving 12 or 13 August as the date. On 10 September 2026, DeepSeek released V4.1-Flash, which added native visual understanding and is reported to have about 552 billion parameters. The V4 models support a context window of one million tokens.

Models and technology

Mixture of experts

Most DeepSeek flagship models use a mixture-of-experts (MoE) architecture. Instead of activating every parameter for each token, a learned router selects a small subset of specialized sub-networks. R1 has 671 billion total parameters but activates about 37 billion per forward pass, a ratio of roughly $37/671 \approx 5.5\%$. This design lets the model draw on a large parameter store while keeping inference compute closer to that of a much smaller dense model.

Attention and numerical efficiency

DeepSeek has published several efficiency techniques. Multi-head latent attention compresses the key-value cache used during attention computation, which reduces memory use during long-context generation. The company also described training with mixed precision, performing most compute-intensive operations in 8-bit floating point (FP8) while keeping some operations in higher precision, to reduce cost on GPU hardware. DeepSeek also reported multi-token prediction as a training objective.

Reinforcement learning for reasoning

R1 was built on the V3 base model and trained with large-scale reinforcement learning to develop chain-of-thought reasoning. According to the R1 technical report, an early variant trained with reinforcement learning alone without supervised fine-tuning showed strong reasoning but produced output that was hard to read, so the final model used a small amount of supervised data before the reinforcement stage. The approach reduced dependence on large volumes of human-labeled reasoning traces, which many other labs had relied on.

Cost and efficiency claims

DeepSeek's cost figures are frequently cited but are easy to misread. The often-quoted figure of roughly 5.6 million US dollars refers to the estimated cost of the final training run of V3, based on rented GPU time, using about 2,000 Nvidia H800 chips. It does not include prior research, failed experiments, data acquisition, or the company's total infrastructure investment, which analysts have described as far larger. The reasoning-stage training cost reported for R1 in a peer-reviewed paper in Nature is much lower, about 294,000 US dollars on 512 H800 GPUs over roughly 80 hours. Export restrictions on advanced chips to China shaped the company's emphasis on hardware efficiency, and this constraint is frequently cited as a reason for its algorithmic work.

Open-weight release and licensing

DeepSeek's main models have generally been released under the MIT license, which permits commercial use, modification, and distillation. Weights are published on platforms such as Hugging Face, and the company's models have accumulated large download counts; one third-party tally put the total above 370 million across more than a hundred public models as of September 2026. The company has also offered API access at prices generally below those of major Western providers, and a free web and mobile chatbot.

Openness has limits. DeepSeek has released model weights and technical reports but has not released its full training code or training datasets. Independent projects, such as the Hugging Face-led Open-R1 effort, have attempted to reproduce the training pipeline using public information, and these efforts depend on details the company has not disclosed.

Organization

DeepSeek is a privately held company and is funded by High-Flyer, which Liang Wenfeng leads as well. Liang serves as CEO of DeepSeek. Third-party sources reported about 160 employees in 2025. The company has described an organizational style with few formal hierarchies, in which teams work on problems they choose and resources are concentrated when a project shows results. Public information about the company's finances, revenue, and staffing remains limited.

Reception and controversies

DeepSeek's models have been praised for technical efficiency and for their open licensing, and they have been widely adopted by developers, researchers, and companies seeking lower-cost alternatives. Benchmark results released by the company are usually confirmed in part by independent tests, but results vary by task, and some evaluations have found that its models lag behind the leading proprietary systems on certain tasks.

The company's rise has also raised policy and security questions. Governments and organizations in several countries have restricted or reviewed the use of DeepSeek applications on government devices, citing concerns about data handling and the jurisdiction in which the service operates. Debates also concern whether the cost of frontier AI is as low as claimed, how far export controls can be effective, and what level of openness is appropriate for models of this capability. Reports on these questions differ in emphasis, and the company has generally addressed them through technical publications rather than detailed public responses on specific allegations.

References

1. DeepSeek History (2026): Liang Wenfeng, R1, V4 and V4.1taskade.com
2. DeepSeeken.wikipedia.org
3. 'Pride of His Hometown': Who Is DeepSeek Founder Liang Wenfeng? What to Know About the 40-Year Old Billionaireentrepreneur.com
4. Liang Wenfeng（Founder and Chief Executive Officer of DeepSeek）_Baiduwikibaike.baidu.com
5. Who is the Founder of DeepSeek AI, Liang Wenfeng?seo.ai
6. DeepSeek V4 Targets Coding Dominance with Mid-Februaryintrol.com
7. DeepSeek officially launches V4-Pro AI model in August 2026tech.yahoo.com
8. DeepSeek Models in October 2026: V4.1-Flash and V4-Proplatform.teamai.com
9. DeepSeek (chatbot)en.wikipedia.org
10. DeepSeek V4: Models, Specs, Pricing, and What's Current (2026)yottalabs.ai
11. AI Model Release Trackerevertune.ai
12. Deepseek Release Notes - September 2026 Latest Updates - Releasebotreleasebot.io
13. DeepSeek Statistics 2026: Users, Models, Costs and Adoption - Second Talentsecondtalent.com
14. 2025 in artificial intelligenceen.wikipedia.org
15. DeepSeek-R1: The Open-Source Reasoning Modelsitepoint.com
16. Comprehensive Analysis of Transparency and Accessibility of ChatGPT, DeepSeek, And other SoTA Large Language Modelsarxiv.org
17. How to Run DeepSeek R1 with Ollama (October 2026)thundercompute.com
18. DeepSeek's Low Inference Cost Explained: MoE & Strategyintuitionlabs.ai
19. Open-R1: a fully open reproduction of DeepSeek-R1huggingface.co
20. DeepSeek is open-access and the next AI disrupter for radiologyncbi.nlm.nih.gov
21. DeepSeek R1: What It Is, How It Works, and Why It Mattersgeneralcompute.com
22. DeepSeek R1: Reasoning, Pricing, and Coding Use - Verdent Guidesverdent.ai