DeepSeek

DeepSeek is a Chinese AI company from Hangzhou, and the name of its chatbot and AI models. You can chat with it in an app, or download the models and run them on your own computers, also for commercial use. For a business, that is the main choice: DeepSeek's own services store your data in China, while a copy of the model on your own server keeps it in-house.

Which models has DeepSeek released?

DeepSeek publishes most of its models as open weights: anyone can download the trained model. The best-known releases:

• DeepSeek-V3 (December 2024): a large language model of 671 billion parameters, of which 37 billion are used per token. Its full training took 2.788 million GPU hours on Nvidia H800 chips, or $5.576 million at $2 per hour, not counting earlier research (technical report).
• DeepSeek-R1 (20 January 2025): a reasoning model that works through a problem step by step before it answers, comparable to OpenAI's o1 on maths, code and reasoning according to DeepSeek. R1 came under the MIT licence, which allows commercial use and distillation, with six smaller versions of 1.5 to 70 billion parameters, built on Qwen and Llama models.
• DeepSeek-V4 (24 April 2026, as a preview): V4-Pro has 1.6 trillion parameters with 49 billion active, V4-Flash 284 billion with 13 billion active, both with a context window of 1 million tokens. In September 2026 V4.1-Flash followed, with an encoder-decoder design.

Why did R1 draw so much attention?

With R1 and its precursor R1-Zero, DeepSeek showed that reasoning can be taught largely through reinforcement learning with GRPO: the model is rewarded when its final answer to a maths or programming problem is correct, and learns by itself to verify and reflect on its steps (DeepSeek-R1 paper). According to Hugging Face (March 2026), China's AI ecosystem "steered heavily into open source" after R1. In the year before the report, Chinese models took 41% of downloads on the platform, more than any other country, and DeepSeek-R1 became the most-liked model (see Why is China ahead in open-source AI?).

How does it work?

Two techniques from DeepSeek-V2 (May 2024) are also in V3, the base of R1:

• Mixture of experts: the model consists of many sub-networks, and each token uses only a few of them.
• Multi-head latent attention (MLA) compresses the KV cache, the memory a model keeps of the text so far. V2 needed 93.3% less KV cache than DeepSeek's earlier 67B model.

What should a business watch out for?

• Data: according to its privacy policy, DeepSeek stores personal data in China and uses it to train its models. In January 2025 the Dutch data protection authority AP advised users to think twice before entering personal or sensitive data, and warned that uploading other people's data can make you liable.
• Security and censorship: in September 2025 CAISI, the AI evaluation centre of the US standards institute NIST, found that agents built on R1-0528 were on average 12 times more likely than US frontier models to follow malicious instructions (see prompt injection), and that the models echoed four times as many misleading Chinese Communist Party narratives as US models.
• Claims: by DeepSeek's figures, V4-Pro is about as capable as Claude Opus 4.6 and GPT-5.4. In CAISI's tests (April 2026), it performed like GPT-5, released some eight months earlier, and was more cost-efficient than GPT-5.4 mini on five of seven benchmarks.
• Licence: the distilled R1 versions built on Llama fall under Meta's Llama licence.

Example

Suppose an accounting firm wants AI to summarise client letters, but client data may not leave the office. The firm can download one of the smaller R1 versions and run it on its own server. The data stays in-house; the trade-off is that a small model scores lower on benchmarks than DeepSeek's large models.