DeepSeek-V4 Preview Released

DeepSeek-V4 Preview has been officially released and open-sourced. This update focuses on providing a context length of 1 million tokens at a very economical cost.

The model is available in two versions:

• DeepSeek-V4-Pro: With 1.6T total parameters and 49B active parameters, it boasts performance competing with the world's best closed-source models. It achieved SOTA on agentic coding benchmarks and shows strong reasoning capabilities in math and STEM fields.
• DeepSeek-V4-Flash: With 284B total parameters and 13B active parameters, it maintains performance close to the Pro model while offering fast response times and high cost efficiency.

Technically, it introduces a new attention mechanism combining token-level compression with DSA (DeepSeek Sparse Attention). This dramatically reduces computation and memory costs while supporting a 1 million token long context as a default specification.

The API is currently available and supports both OpenAI ChatCompletions and Anthropic API formats. It also supports both Thinking/Non-Thinking modes, providing an optimal reasoning environment for each use case.