DeepSeek-V4-Pro Review

The larger half of DeepSeek's V4 release, published on 13 August 2026 after V4-Flash arrived on 31 July. It is a sparse mixture-of-experts model with 1.6 trillion total parameters and roughly 49 billion active on any given forward pass, which is what lets a model this large run at a cost per token nothing like its size would suggest. Context extends to a million tokens, and the model supports several distinct reasoning modes rather than a single fixed one.

Crucially it ships under the MIT license with open weights, which puts a frontier-scale system inside organizations that cannot send data to a hosted API at all, and makes self-hosting a genuine option rather than a technicality. It is text-only, with no vision or audio input. The smaller V4-Flash sits at 284 billion parameters for teams without the hardware to serve the Pro weights. This grade is a provisional placeholder pending hands-on testing.