DeepSeek’s V4 preview made million-token context standard
The April release introduced Pro and Flash variants, open weights and updated API model names.
THE STORY IN THREE POINTS
- The April preview introduced Pro and Flash.
- Both variants offered a million-token context window.
- The announcement linked open model weights.
The announcement
DeepSeek introduced the V4 preview on April 24 with Pro and Flash variants and open weights. The announcement lists 1.6 trillion total parameters for Pro and 284 billion for Flash, with smaller active parameter counts. It made million-token context standard across official services and supported thinking and non-thinking modes. These were preview-era specifications; later Flash and Pro updates followed during the summer.
Editorial context
A larger context window can reduce the need to split a long document, but it does not guarantee that every relevant detail will guide the answer. Readers evaluating document work could place important evidence at different positions and ask questions that require reconciling distant passages. The goal is dependable use of context rather than merely accepting it.
Editorial context: what to watch
For open-weight deployments, available memory, inference support and the intended workload all affect practicality. A useful trial would publish the input length, model version and retrieval setup alongside its results. This April milestone helps explain the current product family while preserving the distinction between the original preview and the models available in the latest API.
Sources: DeepSeek: V4 preview release, April 24 ↗
