DeepSeek

DeepSeek (officially Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., Chinese: 杭州深度求索人工智能基础技术研究有限公司; literally "Deep Exploration") is a Chinese artificial intelligence research company that develops large language models and related AI technologies. Founded in 2023 in Hangzhou by entrepreneur Liang Wenfeng and backed by his quantitative hedge fund High-Flyer, the company rose to international prominence in January 2025 with the release of DeepSeek-R1, a reasoning model whose performance rivaled leading Western frontier models while reportedly requiring only a fraction of the training cost. DeepSeek is widely known for releasing its models as open weights, for its innovations in efficient model architecture and training, and for the significant disruptions its releases caused in global technology markets and AI policy debates.

Background and Founding

DeepSeek's origins lie in High-Flyer (幻方量化), a quantitative investment firm co-founded by Liang Wenfeng in 2015. High-Flyer became one of China's largest quantitative hedge funds, managing assets measured in billions of dollars and relying heavily on machine learning for trading strategies. Beginning around 2019, High-Flyer invested heavily in computing infrastructure, eventually accumulating a cluster of approximately 10,000 NVIDIA A100 graphics processing units (GPUs) for AI research and trading applications.

In 2023, as generative AI attracted global attention following the release of ChatGPT, Liang decided to pursue artificial general intelligence (AGI) research directly. DeepSeek was established as an independent company, funded primarily by High-Flyer, with Liang serving as its head. The company differentiated itself from many Chinese technology firms by positioning itself as a fundamental research organization rather than a product-oriented enterprise. Liang stated that DeepSeek's mission was to pursue AGI and that the company intended to embrace the open-source path, viewing the sharing of model weights and technical reports as a contribution to the global research community rather than a competitive liability.

Historical Development

DeepSeek's early releases focused on specialized domains. In November 2023, the company introduced DeepSeek Coder, a series of code-focused language models trained from scratch on large corpora of programming data. This was followed by DeepSeek LLM, a general-purpose language model family released in late 2023 and early 2024, whose 67-billion-parameter version demonstrated that a small, newly formed team could train competitive large models. In February 2024, DeepSeek released DeepSeekMath, which introduced the Group Relative Policy Optimization (GRPO) reinforcement learning algorithm that would later prove central to the company's reasoning models.

A significant technical milestone arrived in May 2024 with DeepSeek-V2, a 236-billion-parameter Mixture-of-Experts model that introduced Multi-head Latent Attention (MLA), a mechanism that drastically reduces memory requirements during inference. DeepSeek-V2 attracted attention not only for its performance but also for its aggressive API pricing, which triggered a price war among Chinese AI providers. Successive releases in 2024 included DeepSeek-Coder-V2, the vision-language model DeepSeek-VL, and the multimodal Janus series.

In December 2024, DeepSeek released DeepSeek-V3, a 671-billion-parameter Mixture-of-Experts model that activates only 37 billion parameters per token. The accompanying technical report stated that the full training run consumed approximately 2.788 million GPU hours, at an estimated cost of around 5.58 million U.S. dollars using NVIDIA H800 chips—the export-compliant processors available to Chinese firms under U.S. semiconductor restrictions. This figure, though excluding prior research, infrastructure, and personnel costs, was dramatically lower than the reported training expenditures of comparable frontier models and intensified international debate about the efficiency of AI development.

On January 20, 2025, DeepSeek released DeepSeek-R1, a reasoning model trained largely through reinforcement learning to produce extended chains of thought before answering. Benchmark results placed R1's performance at or near the level of OpenAI's o1 model on mathematics, coding, and scientific reasoning tasks. The company simultaneously published the model weights under an MIT license, along with a technical paper detailing its training methodology. Within days, the DeepSeek mobile application became the most downloaded free app on Apple's App Store in the United States, and on January 27, 2025, NVIDIA's shares fell approximately 17 percent, erasing roughly 600 billion dollars in market capitalization in a single day—the largest one-day loss for a single company in stock market history at that time. Venture capitalist Marc Andreessen described the release as "AI's Sputnik moment."

Following this breakthrough, DeepSeek continued iterating on its models in 2025, releasing updated versions such as DeepSeek-R1-0528 and the DeepSeek-V3.1 and V3.2 series, which further improved reasoning capabilities, agent-oriented performance, and inference efficiency, including sparse attention techniques.

Models and Products

DeepSeek's model families include:

• DeepSeek LLM: General-purpose dense language models (7B and 67B parameters) that established the company's pre-training and alignment foundations.
• DeepSeek Coder / DeepSeek-Coder-V2: Code-specialized models trained on multilingual programming data, among the earliest open models competitive with closed coding assistants.
• DeepSeekMath: A mathematics-focused model that introduced the GRPO algorithm.
• DeepSeek-V2 and DeepSeek-V2.5: Mixture-of-Experts models featuring Multi-head Latent Attention, notable for low inference cost and API pricing.
• DeepSeek-V3: A 671B-parameter MoE model (37B activated per token) trained with FP8 precision and auxiliary-loss-free load balancing, establishing a new efficiency benchmark for open models.
• DeepSeek-R1 and R1 updates: Reasoning models trained via large-scale reinforcement learning, with distilled variants based on smaller open models (Qwen and Llama) that made reasoning capabilities accessible on consumer hardware.
• DeepSeek-Prover: Models for formal theorem proving in the Lean programming language.
• DeepSeek-VL and Janus / Janus-Pro: Vision-language and unified multimodal understanding-generation models.

The DeepSeek assistant, available as a web service and mobile application, provides conversational access to the company's flagship models and accounts for much of the brand's public visibility.

Technical Characteristics

DeepSeek's research is distinguished by several architectural and engineering innovations:

1. Mixture-of-Experts (MoE) architectures: DeepSeekMoE employs fine-grained expert segmentation combined with shared experts, improving parameter efficiency compared with conventional MoE designs.
2. Multi-head Latent Attention (MLA): By compressing key-value caches into a latent space, MLA substantially reduces memory consumption during inference, lowering serving costs.
3. Multi-Token Prediction (MTP): Used in DeepSeek-V3's training to improve data efficiency and enable speculative decoding at inference time.
4. Reinforcement learning for reasoning: DeepSeek-R1 demonstrated that chain-of-thought reasoning could emerge from rule-based reinforcement learning with verifiable rewards (GRPO), without the need for extensive supervised fine-tuning on human-labeled reasoning traces.
5. Efficient training engineering: Use of FP8 mixed-precision training, custom parallelism strategies, and optimized GPU communication enabled training on hardware subject to export limitations.

Features and Business Model

DeepSeek's most distinctive characteristics include its commitment to open-source releases, its unusually low pricing, and its research-driven organizational culture. Model weights are published on platforms such as Hugging Face under permissive licenses, allowing developers worldwide to download, modify, and deploy them. The company's API prices have been among the lowest in the industry, sometimes undercutting comparable offerings by an order of magnitude, which accelerated adoption by developers and third-party AI providers.

Unlike large technology conglomerates, DeepSeek operates with a comparatively small team of researchers, many recruited directly from top Chinese universities such as Peking University and Tsinghua University, and it emphasizes young talent, flat management, and long-term research over near-term monetization. Funding from High-Flyer provides financial independence, allowing the company to prioritize fundamental AGI research. Liang Wenfeng has framed DeepSeek's approach as that of a contributor to, rather than a follower within, the global AI innovation ecosystem.

Impact and Significance

DeepSeek's emergence had far-reaching consequences across several domains:

• Economic and market impact: The January 2025 market reaction raised questions about the valuations of AI infrastructure companies and demonstrated that frontier-level AI could be developed at substantially lower cost than prevailing assumptions suggested.
• Technology policy: The success of DeepSeek under U.S. export controls on advanced chips fueled debate in the United States and allied countries about the effectiveness of semiconductor restrictions and the pace of Chinese AI development.
• Open-source ecosystem: By releasing frontier-level models with open weights and detailed technical reports, DeepSeek accelerated the diffusion of reasoning-model techniques and strengthened the position of open model ecosystems globally.
• Industry competition: Its pricing and open releases pressured both Chinese and international AI providers on cost and compelled broader reconsideration of closed, proprietary development strategies.

Reception and Controversies

DeepSeek's rapid rise also generated scrutiny. Several governments and public institutions restricted or banned the DeepSeek application on official devices over concerns about data storage, potential access by Chinese state authorities, and cybersecurity risks; such measures were adopted in jurisdictions including Australia, South Korea, Italy, and parts of the United States. Critics also noted that DeepSeek's models apply content filtering consistent with Chinese regulations, declining to answer politically sensitive questions. Additionally, some industry figures questioned the completeness of reported training-cost figures, and OpenAI alleged in early 2025 that DeepSeek may have used outputs from its models, potentially in violation of terms of service—an accusation that itself sparked debate about the industry's standard data practices. DeepSeek and its supporters have responded that its innovations are independently documented in peer-scrutinized technical reports and reproducible through its open releases.

Despite these controversies, DeepSeek is broadly regarded as one of the most influential AI research organizations of the mid-2020s, having demonstrated that architectural innovation and engineering efficiency can substantially lower the barriers to frontier-scale artificial intelligence development.