DeepSeek has released the stable version of its V4-Flash API, and the focus is squarely on coding, tool calling, and long-horizon task execution. The company says the new model handily beats the previous V4-Pro-Preview across several agent benchmarks.
The updated model natively supports the Responses API and has been adapted for OpenAI's Codex. The architecture and parameter count haven't changed — this is essentially a retraining pass.
Only the API is getting the upgrade, though. The DeepSeek app and web client remain unchanged. And if you're waiting for the full V4-Pro stable release, you'll wait a little longer: the company says it will land soon.
DeepSeek