- Sources: primary, discussion
- Summary: DeepSeek's change log dated 2026-07-31 puts the V4-Flash API into public beta, adds the native Responses API format, and adds a Codex adaptation. V4-Flash-0731 keeps the architecture and size of V4-Flash-Preview and was re-post-trained only, and the V4-Pro API along with the app and web models are unchanged. The code-agent figures are stated as run under DeepSeek Harness minimal mode at max effort with top_p 0.95 and temperature 1.0, the harness is described as not yet released, and DSBench-FullStack and DSBench-Hard are named as internal test sets, so those numbers are attributable but not reproducible today.
- Why it matters: The reproducible part of the release is the interface work, because the agent harness behind the reported scores is unreleased and two of the nine test sets are internal.
- Follow-up: Track whether DeepSeek releases the harness and whether the internal test sets are published.
send feedback on this story