The headline here isn't a bigger model or a new architecture — it's a retrained one. DeepSeek dropped the public beta of its V4-Flash API on July 31 under the build tag 0731, and the same 284B-parameter mixture (13B active per token, one-million-token context, MIT license) now outscores the company's own flagship preview on all nine agent and coding benchmarks it reported. Terminal Bench 2.1 climbs to 82.7 from the preview's 61.8; DeepSWE jumps to 54.4 from a near-useless 7.3. If you already call deepseek-v4-flash, you have nothing to do — same endpoint, same key, same model name, and the upgrade is silent. Worth knowing the new build speaks the Responses API format natively and is Codex-compatible, so wiring it into an existing agent loop is close to free. The interesting part is what it says about method over scale: these gains came entirely from post-training, not a redesign.