DeepSeek V4 preview brings a million-token context to two model tiers

DeepSeek released V4 Preview on 24 April in Pro and Flash variants. The company lists 1.6 trillion total parameters with 49 billion active for Pro, compared with 284 billion total and 13 billion active for Flash. Both support a one-million-token context window and thinking or non-thinking modes.

The release includes open weights, web access and updated API model names. DeepSeek says the design combines token compression with sparse attention to handle longer contexts more efficiently. A million-token limit specifies capacity, not a promise that every detail in that input will be used correctly. Developers also need to choose between hosting the weights themselves and using the company’s API, which have different hardware, operating and data-handling implications.