This afternoon (the 31st), DeepSeek announced through an API documentation release log that the official DeepSeek-V4-Flash API has launched its public beta.
According to the announcement, DeepSeek-V4-Flash’s Agent capabilities have been significantly enhanced, far surpassing V4-Pro-Preview in benchmark tests:
Terminal Bench 2.1: 82.7
NL2Repo: 54.2
Cybergym: 76.7
DeepSWE: 54.4
Toolathlon verified: 70.3
Agent Last Exam: 25.2
Automation Bench (Public): 25.1
DSBench-FullStack: 68.7
DSBench-Hard: 59.6
It is reported that the official V4-Flash version natively supports the Responses API format and has been specifically adapted for Codex. The model architecture and size of DeepSeek-V4-Flash-0731 remain consistent with DeepSeek-V4-Flash-preview, with only additional post-training performed. For detailed configuration instructions, see the documentation link: Click here
The official announcement also stated that this update only upgrades the DeepSeek-V4-Flash API interface. The DeepSeek-V4-Pro API and the models on the APP / WEB platforms remain unchanged, and the official version of DeepSeek-V4-Pro will be released as soon as possible.