DeepSeek is a Chinese artificial intelligence company based in Hangzhou and the name of its family of large language models, best known for the general-purpose DeepSeek-V3 and the reasoning model DeepSeek-R1. Among AI models & assistants, DeepSeek stands out for releasing open-weight models that approach the performance of leading proprietary systems at a substantially lower reported training and inference cost.
The company was founded in 2023 with funding from the quantitative hedge fund High-Flyer, led by Liang Wenfeng. It also operates a free chat assistant on the web and in mobile apps, plus a paid developer API. The release of DeepSeek-R1 in January 2025 prompted wide debate about the cost of frontier AI and about export controls on advanced chips.
The DeepSeek model family and its releases
DeepSeek has released models at a rapid, incremental pace, with each generation introducing an architectural or training change that later releases reused. The table below lists the main milestones; parameter counts are those published by the company.
| Model | Released | Size | Main contribution |
|---|---|---|---|
| DeepSeek LLM | Late 2023 | 7B and 67B, dense | First general open models |
| DeepSeek-Coder | Late 2023 | Up to 33B, dense | Code-specialized training |
| DeepSeek-V2 | May 2024 | 236B total, 21B active | Multi-head latent attention, sparse experts |
| DeepSeek-V3 | December 2024 | 671B total, 37B active | FP8 training, multi-token prediction |
| DeepSeek-R1 | January 2025 | 671B total, 37B active | Reasoning learned through reinforcement |
| DeepSeek-V3.1 and V3.2 | 2025 | 671B class | Hybrid thinking modes, sparse attention |
Alongside the flagship models, the company published distilled R1 variants from 1.5B to 70B parameters, built by fine-tuning Qwen and Llama base models on reasoning traces generated by DeepSeek-R1. These smaller models run on a single workstation GPU and made local reasoning models widely accessible.
Architecture and efficiency techniques
DeepSeek’s efficiency comes mainly from a mixture-of-experts (MoE) architecture, in which only a small subset of parameters is used for each token. In DeepSeek-V3, roughly 37 billion of 671 billion parameters are active per token, so inference compute resembles that of a much smaller dense model while total knowledge capacity stays large.
Several further techniques reduce memory and compute. The following table summarizes the most cited ones and what each achieves.
| Technique | What it does | Practical effect |
|---|---|---|
| DeepSeekMoE | Many fine-grained experts plus shared experts | Better specialization per active parameter |
| Multi-head latent attention | Compresses keys and values into a latent vector | Much smaller key-value cache |
| Auxiliary-loss-free load balancing | Adjusts routing bias instead of adding a penalty | Even expert use without hurting quality |
| Multi-token prediction | Trains the model to predict several future tokens | Denser training signal, faster decoding |
| FP8 mixed precision | Runs most training math in 8-bit floating point | Lower memory and faster training |
| DeepSeek Sparse Attention | Attends only to selected earlier tokens | Cheaper long-context processing |
DeepSeek reported that the final training run of DeepSeek-V3 used about 2.8 million H800 GPU hours, which it priced at roughly 5.6 million US dollars. This figure excludes research, failed experiments, data work and hardware ownership, so it is not the total cost of building the model, a distinction often lost in public discussion.
How DeepSeek-R1 learned to reason
DeepSeek-R1 is a reasoning model that produces a long chain of thought before giving its final answer, in the same class as the OpenAI o-series. Its notable feature is the training recipe, which relied heavily on large-scale reinforcement learning with automatically checkable rewards.
The precursor, DeepSeek-R1-Zero, was trained with reinforcement learning directly on the V3 base model, without supervised examples of reasoning. Rewards came from rule-based checks, such as whether a math answer matched or code passed tests. The algorithm, Group Relative Policy Optimization (GRPO), compares several sampled answers to the same prompt and avoids a separate critic model, which saves memory.
R1-Zero showed emergent behaviors such as self-verification, but its outputs were often hard to read and mixed languages. The final R1 therefore added a small cold-start supervised stage, further reinforcement learning and a broader fine-tuning round.
| Stage | Data or signal | Purpose |
|---|---|---|
| Cold-start fine-tuning | Thousands of curated reasoning examples | Readable chain-of-thought format |
| Reasoning reinforcement | Rule-based correctness rewards | Stronger math, code and logic |
| Rejection sampling and tuning | Filtered model outputs plus general data | Broader skills beyond reasoning |
| Final reinforcement | Correctness and preference rewards | Helpfulness and harmlessness |
DeepSeek compared with other model families
DeepSeek is usually compared with other open-weight families such as Llama and Qwen, and with proprietary assistants such as Claude, Gemini and GPT-5. Its distinctive position combines frontier-level reasoning scores, a permissive license and low API prices.
| Aspect | DeepSeek | Typical proprietary model | Typical open-weight model |
|---|---|---|---|
| Weights | Published, MIT license for R1 | Not published | Published, license varies |
| Architecture | Large sparse MoE | Mostly undisclosed | Dense or MoE |
| Self-hosting | Possible but hardware-heavy | Not possible | Often practical |
| API pricing | Low per million tokens | Medium to high | Depends on host |
| Modalities | Mainly text | Text, image, audio | Mostly text |
Running the full 671B model locally requires several hundred gigabytes of accelerator memory, even with quantization, so most organizations either use a cloud host or choose a distilled variant.
Limitations, censorship and data concerns
DeepSeek models carry content restrictions shaped by Chinese regulation. The hosted assistant declines or deflects questions on politically sensitive topics, and similar behavior appears to varying degrees in the open weights. Organizations should test for these biases before deploying the models in public-facing products.
- Data residency: the hosted app and API process and store data on servers in China, which has led several governments and agencies to restrict use on official devices.
- Reasoning overhead: R1 generates long thinking traces, increasing latency and token cost for simple tasks.
- Safety testing gaps: independent red-team evaluations have reported weaker refusal of harmful requests than leading proprietary models.
- Limited multimodality: the main chat models focus on text, while vision support has come through separate releases.
Self-hosting the open weights removes the data residency concern, because no prompts leave the organization, but it does not remove the trained-in content behavior.
The idea in brief
DeepSeek is a Chinese company that makes AI chatbots and the models behind them. Its models can write, answer questions, solve math problems and write code. Many people use its free chat app.
What made DeepSeek famous is that it reached near top-level quality at a lower cost. It also gave away the model files so anyone can download and run them. Think of a carmaker that builds a car almost as fast as the best, sells it cheaply and publishes the engine design. The catch is that the hosted service keeps data in China and avoids some political topics.
A practical scenario
A mid-sized engineering firm wants an internal assistant that helps staff check calculations and summarize technical reports. Its legal team will not allow confidential documents to be sent to an external service. The IT department therefore deploys a distilled DeepSeek-R1 model with 32 billion parameters on two in-house GPU servers, quantized to 4-bit precision.
In a pilot with 40 engineers, the model handles unit conversions and step-by-step load calculations well, though answers take 20 to 60 seconds because of long reasoning traces. The team adds a fast non-reasoning model for simple questions and routes only calculation tasks to R1. The result is useful reasoning support with no data leaving the building, at a hardware cost the firm could justify after three months of use.