What is DeepSeek

DeepSeek is both a family of large language models and the AI assistant built on them, developed by Hangzhou-based DeepSeek. The company focuses on research into foundation models for general AI, and stands out for publishing its main models openly. From DeepSeek-V3 and DeepSeek-R1 to the current V4 series, the weights are released on Hugging Face and ModelScope, so developers can download them to run locally or study.

For everyday users, DeepSeek is a chat assistant with a “DeepThink” reasoning mode and web search. For developers, it offers an API whose format is compatible with both OpenAI and Anthropic, so many existing tools can connect by changing a few settings.

The change log in the API docs gives the main milestones: DeepSeek-V3 in December 2024, the DeepSeek-R1 reasoning model and the official app in January 2025, the open DeepSeek-V4 preview in April 2026, and later the V4-Pro release and V4.1-Flash.

Much of the attention DeepSeek gets comes from pairing openness with usability. Anyone can use the latest models free on the web and in the app, while the published weights and technical reports let researchers see the design trade-offs and let companies run the same models on their own servers. Many teams validate an idea on the official API first and move to private deployment later without worrying about a drop in quality from switching models.

On the technical side, DeepSeek says V4 introduces a new attention mechanism that compresses along the token dimension, combined with DSA (DeepSeek Sparse Attention). This supports very long contexts while sharply reducing compute and memory needs, which is what makes 1M-token context practical as a default across its services.

Key features

Chat. Ask questions on the web or in the app for writing, translation, Q&A and code explanations.

DeepThink. When enabled, the model reasons before answering, which helps with maths, logic and programming. In the API, the V4 models support both thinking and non-thinking modes, and the reasoning_effort parameter sets how hard the model thinks.

Web search. The assistant can search for current information and use it in its answer, working around the cut-off of its training data.

Million-token context. With the V4 preview, DeepSeek announced that a 1M-token context is standard across all its official services, enough for very long documents or a large codebase in one pass.

Two model tiers. V4 comes as Pro and Flash. Pro is stronger; Flash has fewer parameters and activations, so it is faster and cheaper. DeepSeek says Flash is close to Pro on simple agent tasks but behind on hard ones.

Tuned for agents. DeepSeek says V4 was adapted for popular agent products including Claude Code, OpenCode, OpenClaw and CodeBuddy, and the API docs include integration guides for them. It also offers DeepSeek Harness, a developer preview for agent harness builders, with a desktop client.

Open weights. Weights and technical reports for V4 and earlier models are available on Hugging Face and ModelScope for private deployment and research.

How to use it

As a user:

  1. Go to chat.deepseek.com or install the DeepSeek app.
  2. Sign in and ask. Turn on DeepThink for careful reasoning and search when you need current information.
  3. Upload documents to have them summarised or questioned.

As a developer:

  1. Create an API key on the DeepSeek platform.
  2. With the OpenAI SDK, set base_url to https://api.deepseek.com. With the Anthropic SDK, use https://api.deepseek.com/anthropic.
  3. Set model to a name currently listed in the docs (for example deepseek-v4-pro) and send your request.

Model names change between releases. DeepSeek announced that the legacy names deepseek-chat and deepseek-reasoner would stop working on 24 July 2026, so check the API docs before integrating.

If you already use a coding agent such as Claude Code or OpenCode, you can make DeepSeek its backend without writing code by following the Agent Integrations section of the API docs. For complex agent tasks, DeepSeek recommends thinking mode with the effort set to max.

Use cases

  • Learning and Q&A: explaining concepts and helping with maths and programming problems.
  • Writing and translation: drafting, polishing and translating between Chinese and English.
  • Long documents: reading whole contracts, reports or codebases thanks to the long context.
  • Programming: chat-based coding, or DeepSeek as the backend model for agents like Claude Code.
  • Private deployment: run the open weights on your own hardware when data must stay in-house.

FAQ

Is DeepSeek open source? The weights of its main models are published on Hugging Face and ModelScope. Check the licence file in each model repository for the exact terms.

When should I turn on DeepThink? Use it for multi-step problems such as proofs, logic and complex code. For simple questions, translation or casual chat, leaving it off gives faster replies.

Why did my old integration start failing? The model name has probably changed. Check the Models & Pricing page in the API docs and update the model parameter.

How is the API priced? By tokens, with different prices per model. See the official Models & Pricing page for current rates.

Official website