Pros

• Similar capabilities to high-end subscriber-only US models.
• Shows thinking steps when using R1 model.
• Also available as local model (open source).

Cons

• Trains actively on user inputs, which cannot be disabled.
• Text-only, no multimodal capabilities.
   

Whilst DeepSeek has been around for a couple of years already with impressive coding and language models, the release of their V3 and R1 models to the public late 2024 / early 2025 grabbed media headlines. This China-based company managed to build a language model similar in quality to the US market leader (OpenAI), despite restrictions on access to advanced chips and a smaller budget. Contrary to OpenAI, DeepSeek released their model to the open source community to allow people to use it locally, as well as via their app and website. This included their R1 model which best compares to OpenAI’s o1 model, which OpenAI has kept as subscriber-only model up to that point.

DeepSeek’s R1 model stands apart from its competition in both the accessibility and transparency. During the operation of the model the user can ‘see’ the model think, giving an insight into the ‘reasoning’ applied by the model.

The biggest criticism of DeepSeek lies with its policy towards privacy and data security. As a user you are unable to control what data you share with DeepSeek when using the model. In the company’s privacy statement it is clearly mentioned that inputs will be used for further feature development. This presents a significant risk when using the model via the app or web interface. Usage locally is therefore recommended if usage of this model is considered .