On 24 April 2026 DeepSeek released a preview of two new open-weight models, DeepSeek-V4-Pro and DeepSeek-V4-Flash. Both ship with a 1 million token context window under the MIT license on Hugging Face. The release keeps DeepSeek’s flagship line fully open.

What happened

DeepSeek released a preview of DeepSeek-V4 on 24 April 2026. The release includes two models.

DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active. DeepSeek-V4-Flash has 284 billion total parameters with 13 billion active. Both models offer a 1 million token context window, and both are published as open weights under the MIT license on Hugging Face.

Why it matters for builders

The large gap between total and active parameters marks these as /glossary/mixture-of-experts/ designs: only a fraction of the network runs per token, so a 1.6 trillion parameter model activates 49 billion at inference. That keeps compute cost far below what the total size implies.

A 1 million token context window makes DeepSeek-V4 a /glossary/long-context-model/, suited to long documents, large codebases, and extended agent histories. The MIT license permits commercial use and self-hosting without restriction, which matters for teams that need to run models on their own infrastructure. See /tools/deepseek/ for deployment options.

Sources

  1. DeepSeek. “DeepSeek-V4 release” (24 April 2026). https://api-docs.deepseek.com/news/news260424