DeepSeek moved its Flash-tier model out of preview and into general availability under the label V4-Flash-0731, and the numbers behind that release are notable less for a big architecture change than for what a single round of retraining can do. The model keeps the exact same shape as the April preview build -- a 284-billion-parameter mixture-of-experts model that only activates about 13 billion parameters per token -- but DeepSeek redid the post-training stage, and the resulting benchmark jump is large: Terminal-Bench 2.1 score rose from 61.8 to 82.7, and DeepSeek's own agentic coding benchmark, DeepSWE, jumped from roughly 7.3 to 54.4. On a real-world agentic work benchmark (GDPval-AA v2), its Elo score rose from 1189 to 1559. It now beats DeepSeek's own larger Pro-tier model on every one of the nine benchmarks DeepSeek reports, by margins ranging from about 9 to 42 points, which is an unusual result for a "smaller, faster" model to post against its own flagship sibling. Pricing stayed flat at $0.14 per million input tokens, and the release doubles concurrency headroom to 2,500 simultaneous requests versus 500 for the Pro tier -- a detail that matters more for anyone running fleets of coding agents in parallel than any single benchmark number does. The practical takeaway for developers is that post-training quality, not just parameter count, is now doing a lot of the work in closing the gap between "cheap and fast" and "capable enough for agentic coding," and DeepSeek is explicitly positioning the Flash tier as the default for agent workloads rather than a lightweight fallback.