DeepSeek’s smaller model just outperformed its own flagship

DeepSeek released DeepSeek-V4-Flash-0731 as a public beta, a smaller update to its Flash line.

It keeps the same architecture as the earlier Flash preview but adds post-training improvements for agent tasks.

The model is available through DeepSeek’s API, and MIT-licensed weights were published on Hugging Face.

DeepSeek says the update beats its earlier V4-Pro preview on multiple benchmarks.

An independent test reported a lower score than DeepSeek’s claim, underscoring mixed performance signals.