DeepSeek moved its V4-Flash line out of preview on July 31, publishing the "0731" build as the official release under an MIT license, with the full download available on Hugging Face. The headline change isn't a bigger model — the architecture, the 284-billion-parameter mixture-of-experts design, and the roughly 13-billion active-parameter footprint are unchanged from the preview build. What DeepSeek did instead was run another post-training pass specifically aimed at agentic behavior: multi-step tool use, longer-horizon task execution, and tighter adherence to structured output formats that downstream frameworks depend on. The team also added native support for the Responses API pattern that OpenAI popularized and made explicit accommodations for running the model inside Codex-style coding harnesses, which matters if you're wiring this into an existing agent stack rather than building one from scratch. For developers, the practical draw is that V4-Flash-0731 reportedly outperforms DeepSeek's own larger V4-Pro preview on several benchmarks despite activating a fraction of the parameters, meaning cheaper inference without giving up much on coding and tool-use tasks. Because it ships as open weights with a 1-million-token context window, teams that need to self-host for cost, latency, or data-residency reasons now have another serious option alongside Kimi K3 and other open releases from the past few weeks, rather than defaulting to a closed API. The fact that this was a re-post-training exercise rather than a fresh pretrain also signals where the open-weight labs are currently competing hardest: not on raw scale, but on squeezing more reliable agentic behavior out of architectures they already have, which is a cheaper and faster iteration loop than training a new frontier model from zero.