DeepSeek is a Chinese AI lab based in Hangzhou that builds and releases large language models, most of them under open licenses. It became widely known for shipping frontier-class reasoning and coding models at a fraction of the price charged by closed-source competitors, and for publishing technical reports detailed enough that other labs could reproduce the ideas.
The current generation is DeepSeek V4, released in April 2026 as two Mixture-of-Experts tiers: V4 Pro for the hardest reasoning and agentic coding, and V4 Flash for high-throughput, lower-cost work. Both share a 1M-token context window and expose reasoning effort as a request parameter rather than a separate model ID.
DeepSeek also offers an OpenAI-compatible API. That compatibility is the practical reason DeepSeek shows up so often in agent tooling: if a coding agent lets you set a custom base URL and API key, you can usually point it at DeepSeek without patching the tool itself.