What changed
DeepSeek moved V4-Flash out of preview on July 31, 2026 with the 0731 build, an official API release its own change log calls a public beta 1. The build keeps the same architecture and size as V4-Flash-Preview and was only re-post-trained 1. It now supports the Responses API format natively and is adapted for Codex, while V4-Pro and the app and web models stay untouched 1.
The scores ride on a harness you cannot download
DeepSeek reports Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, Cybergym at 76.7, DeepSWE at 54.4 and Toolathlon verified at 70.3 1. Those code-agent scores were measured with DeepSeek Harness minimal mode, which the same note says is still to be released, at the max effort level with topp=0.95 and temperature=1.0 1. Until that harness ships, those scores are a claim I have to take on trust. The baseline is picked too: the comparison is V4-Pro-Preview 1, the preview of the larger sibling 2, not the Flash preview this build replaces.
What I would run
The stack has not moved since April 24, 2026: a 284B MoE with 13B active per token, a 1M-token context window, MIT-licensed weights 2. Those weights are why this one earns an hour where a hosted-only launch would not, since you can pin them the day the hosted build shifts again. In OpenCode it lands as opencode-go/deepseek-v4-flash with two reasoning variants, high and max, selected with --variant 2. Developers Digest, running it as its default, puts max on multi-step agent loops and high on single-file edits and review passes 2.
Impact on your team
If an agent of yours runs on V4-Flash, re-run your eval suite this week: the weights changed behind an API upgrade 1, and the vendor scores came off a harness nobody outside DeepSeek has 1. Start multi-step work on --variant max 2, and keep high for review passes where latency is the real cost 2. What I would not do this week is rewrite a Codex integration onto the native Responses API path 1: it is a public beta 1, and the harness that would let you verify the upgrade has no ship date 1.