If you've been holding out for DeepSeek's polished V4-Pro, keep holding. Today's news is all about V4-Flash, which just got its official API release — and it's bringing meaningful upgrades to coding, tool calling, and long-horizon agent tasks. DeepSeek says its agent benchmark scores now comfortably beat the earlier V4-Pro-Preview.
The model architecture and parameter count haven't changed; this release is essentially a post-training pass. It also natively supports the Responses API and comes with optimizations for Codex.
One catch: this is strictly an API update. The DeepSeek app and web interface aren't getting the new model. And the full V4-Pro? Still no launch date, though DeepSeek says it will be out "as soon as possible."