TheModelDesk

DeepSeek has opened its V4-Flash API in public beta. No calling changes are required: setting the model parameter to deepseek-v4-flash selects the new version.

The release claims significantly improved agent capabilities, with benchmark results described as far above V4-Pro-Preview: Terminal Bench 2.1 at 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, and DSBench-Hard 59.6. For the Code Agent tasks in the public sets, DeepSeek tested with its soon-to-be-released DeepSeek Harness in minimal mode, at max effort level, topp=0.95 and temperature=1.0. DSBench-FullStack and DSBench-Hard are internal DeepSeek test sets for full-stack development and hard coding-agent problems respectively.

V4-Flash natively supports the Responses API format and is specifically adapted for Codex, with configuration details in the documentation. The checkpoint, internally V4-Flash-0731, keeps the same architecture and size as V4-Flash-Preview and differs only through re-post-training.

The update touches only the V4-Flash API: the V4-Pro API and the app/web models are unchanged. DeepSeek says the official V4-Pro release will follow soon. Background from the changelog: since 2026-04-24 the API exposes deepseek-v4-pro and deepseek-v4-flash via both OpenAI ChatCompletions and Anthropic interfaces, and the legacy names deepseek-chat and deepseek-reasoner were scheduled for retirement on 2026-07-24, meanwhile aliasing V4-Flash's non-thinking and thinking modes.