DeepSeek, the Chinese lab that releases its model weights

A low-profile company with high-profile models

DeepSeek is the trading name of Hangzhou DeepSeek Artificial Intelligence Co., Ltd., a company registered in Hangzhou, China. It says little about itself: its website focuses on its models, technical documentation and research papers rather than its corporate story. Its reputation was built on what it released, especially reasoning models made freely available together with detailed technical reports.

Two lines, published weights

DeepSeek models fall into two lines:

• DeepSeek-V: the general-purpose models. The V4 generation comes in a Pro version, the most capable, and a Flash version, faster and cheaper, which can now read images too.
• DeepSeek-R: the reasoning models, which work through intermediate steps before giving a final answer.

DeepSeek’s release policy is one of the most open in the industry: the weights are published on Hugging Face, and the recent models we checked are under the MIT license. Their size, however, calls for professional hardware; most individuals will use the assistant or a hosting provider instead.

How to access DeepSeek

• The DeepSeek assistant on the web and as a mobile app: see the DeepSeek page.
• The DeepSeek API, billed by usage and compatible with the OpenAI and Anthropic formats. Many agentic coding tools accept it without code changes.
• Downloading the weights, to host the models on your own servers or with a cloud provider.

What sets DeepSeek apart

In practice, DeepSeek showed that a lab far less visible than the US giants could release top-tier models under a very permissive license. For developers, that means the freedom to choose where and how the model runs. The trade-off is privacy: using the assistant or the API means your data is processed by a Chinese company. To compare it with Alibaba or Mistral, see the open source models page.