DeepSeek V4 Flash 0731 (Tencent Cloud)

DeepSeek released DeepSeek V4-Flash-0731 on July 31, 2026 as a retrained checkpoint of the 284B-parameter MoE model first previewed in April. The architecture is unchanged, with 13 billion parameters active per forward pass, and the entire delta sits in post-training, including an expanded reinforcement-learning phase targeting instruction-following fidelity, tool-call accuracy under extended context, and multi-step chain-of-thought reasoning. The report says the retrained model beats DeepSeek's own V4-Pro-Preview across nine published agent and coding benchmarks, ships MIT-licensed weights on Hugging Face, and adds native OpenAI Responses API support for Codex CLI users at $0.14 per million input tokens. It flags reproducibility caveats on the headline numbers and notes that the more consequential question is how such cheap output pricing behaves on very long coding loops.

DeepSeek-V4-Flash-0731 keeps the same 284B-total, 13B-active Mixture-of-Experts architecture, 1M context window, and speculative decoding module as the preview it replaces, with the model card stating it kept the same architecture and size and was only re-post-trained. Agent benchmark scores moved sharply, with Terminal Bench 2.1 rising from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4. The analysis notes the model is text-in, text-out with no vision, audio, or video support, and is licensed MIT for commercial use and self-hosting. It cautions that all comparisons target V4-Pro Preview and come from DeepSeek's own evaluation harness, so engineers should benchmark independently before migrating, with the naming acting as a July 31, 2026 date-stamped build rather than a new generation.

DeepSeek moved V4 Flash out of preview on July 31, 2026 as DeepSeek-V4-Flash-0731, keeping the exact architecture and size of the April preview build, a 284B MoE with 13B active per token and a 1M-token context window, and the changelog states it was only re-post-trained. The update was scoped to the Flash API only, while the 1.6T-parameter V4-Pro official release remained pending. Besides re-post-training, the build adds native OpenAI Responses API support and is described as specifically adapted for Codex, letting harnesses built on the Responses format run without an adapter. The article also reports agent benchmark scores passing V4-Pro-Preview and documents how to run the model in the OpenCode coding agent.

Nearly all published DeepSeek-V4-Flash-0731 numbers are vendor-reported, with DeepSeek's own table listing Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, Cybergym at 76.7, and DeepSWE at 54.4, produced with the not-yet-released DeepSeek Harness in minimal mode at maximum reasoning effort. Two of the nine rows rely on internal DeepSeek datasets that outsiders cannot rerun. The one same-day independent signal comes from Artificial Analysis, which reran its nine-benchmark harness on 2026-07-31 and measured an Intelligence Index of 50 versus 40 for the prior Flash, with the agentic GDPval-AA v2 Elo rising from 1189 to 1559. The page also notes DeepSeek was absent from the official Terminal Bench leaderboard, where the top listed pairs scored around 83.8.

DeepSeek published DeepSeek-V4-Flash-0731, the official release of DeepSeek-V4-Flash that supersedes the preview build with substantially enhanced agentic capabilities. It keeps the same model structure as DeepSeek-V4-Flash-DSpark, including an attached speculative decoding module, and reports outperforming DeepSeek-V4-Pro (Preview) on agent benchmarks despite a far smaller activated parameter count. The model card lists vendor results including 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, evaluated in DeepSeek Harness minimal mode at max reasoning effort with temperature 1.0 and top p 0.95. It ships no Jinja chat template, instead providing a dedicated encoding folder for OpenAI-compatible message encoding, and reasoning effort now supports low, high, and max levels.