DeepSeek V4-Flash API enters public beta with agent task upgrades

AI company DeepSeek has launched the official version of its V4-Flash API into public beta. The latest update to the programming interface specifically targets improvements for supporting advanced AI agent tasks.

The model's performance has been evaluated across several benchmarks. DeepSeek reported that the V4-Flash model achieved a score of 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE. These results indicate the model's enhanced capabilities, particularly in handling complex assignments.

The new release includes support for the Responses API and has been adapted for the Codex environment. While the model retains the same structure and size as its preview version, its retraining signifies an upgrade. DeepSeek clarifies that this update exclusively affects the V4-Flash API; the V4-Pro API and models used on the company's app and website remain unchanged.