DeepSeek's stable V4-Flash API is officially out, and it's built for agents that need to get things done. The update focuses on coding, tool calling, and long-task execution, and DeepSeek says it far exceeds the previous V4-Pro-Preview across multiple agent benchmarks.
The new release natively supports the Responses API and is adapted for Codex. Under the hood, the model architecture and parameter count haven't changed — this is mostly a fresh round of post-training.
This launch is API-only, though. The DeepSeek App and web client aren't getting anything new, and the stable version of V4-Pro still hasn't arrived. DeepSeek says it will land as soon as possible.