Definition:
DeepSeek is a Chinese artificial intelligence company that develops models capable of understanding and generating language, working with code and solving reasoning tasks. Its name also identifies its model families and the conversational assistant through which they can be used.
In addition to providing access through an online service and an API, DeepSeek publishes models that can be downloaded and run on users’ own infrastructure, subject to their technical requirements and licensing terms.
Table of contents
DeepSeek as a company, model and assistant
DeepSeek is based in Hangzhou, China. It gained greater international prominence in 2025 through the spread of its models and the release of DeepSeek-R1, designed for reasoning tasks.
It is useful to distinguish three elements that share its name: the company developing the technology, the models processing information and the assistant with which the user converses.
The assistant includes an interface and features that are not necessarily part of the downloadable model. Running a model on a private server does not automatically reproduce all the capabilities of the online service.
How its models work
DeepSeek models generate responses based on their training, the instructions received and the available context. Instructions, known as prompts, specify the task, provide information and define the format of the result.
Their preparation may combine pre-training, supervised fine-tuning and reinforcement learning. It is not correct to describe all DeepSeek models as systems trained exclusively through trial and error: the procedures vary between models and stages.
For reasoning-oriented models, training aims to improve performance on tasks that require several steps, such as mathematical or programming problems. This does not mean that they understand like a person or guarantee that their conclusions are correct.
What it means for its models to be open
DeepSeek publishes code and weights for certain models: the numerical values learned during training that allow the models to run. This availability makes it possible to use them outside the company’s hosted service.
The DeepSeek-R1 repository distributes its code and weights under the MIT licence, which permits commercial use and modifications subject to its terms. Variants derived from other models may carry additional conditions that need to be reviewed separately.
Openness should not be interpreted as access to all the data and processes used during training either. Available code, downloadable weights and full disclosure of training are different things.
Ways to use DeepSeek
Access can be organised in different ways, depending on the intended use and the level of control required:
- Online assistant: allows users to converse with the models through the interfaces offered by the service.
- API: allows their capabilities to be incorporated into custom applications through requests from another program.
- Self-hosted deployment: runs a downloadable model on a compatible computer or server.
- Third-party providers: offer access to DeepSeek models under their own service terms.
The API makes it easier to integrate tasks such as summarising documents, classifying text or assisting with code without directly managing the model’s infrastructure.
Self-hosting offers more technical control, but requires resources and maintenance. Memory and processing requirements depend on the chosen model: not every model can run on any computer.
What DeepSeek is used for
Its models can support writing, translation, document summarisation and information extraction. In programming, they can explain code, suggest solutions and help investigate errors.
Reasoning models are also used to tackle mathematical problems and tasks that require relating several conditions. Their practical usefulness depends on the task, the model and the quality of the information provided.
When integrated into an application, they can be combined with tools or external sources. These connections have to be implemented: installing a model does not automatically give it access to files, databases or services.
Costs, privacy and data control
Downloading a model under a licence that permits its use does not mean that running it is cost-free. Infrastructure, resource consumption, administration and maintenance are part of the deployment.
It is also important to distinguish where information is processed. Using the online assistant or an API involves sending data to the corresponding provider. A self-hosted deployment can keep processing within the chosen infrastructure, provided that the application and its connections are configured accordingly.
Responses may contain errors, biases or unverified information. The availability of code or weights facilitates technical control, but does not by itself guarantee accuracy, privacy or security.