DeepSeek V4 Flash 0731 (Baidu)

DeepSeek released DeepSeek V4-Flash-0731 on July 31, 2026 as a retrained checkpoint of the 284-billion-parameter MoE model first previewed in April, keeping the same architecture with 13 billion active parameters per token. The change is confined to post-training, with an expanded reinforcement-learning stage aimed at instruction-following fidelity, tool-call accuracy under extended context, and multi-step reasoning. The article says the retrained model beats DeepSeek's own V4-Pro-Preview on nine published agent and coding benchmarks, that the weights ship under an MIT license on Hugging Face, and that the API introduces native OpenAI Responses API support for Codex CLI users, with input pricing cited at $0.14 per million tokens. It flags reproducibility asterisks around the headline numbers.

DeepSeek-V4-Flash-0731 is the official, non-preview release of DeepSeek's efficiency-tier V4 model, a 284B-total / 13B-active mixture-of-experts system with a 1M-token context window, an attached speculative decoding module, and MIT licensing for commercial use and self-hosting. The naming is a date stamp for the July 31, 2026 build, and it supersedes the V4-Flash Preview while sitting below the 1.6T-parameter V4-Pro flagship. It handles text only, with no vision, audio, or video support. The analysis notes that architecture did not change: the model card says it keeps the same architecture and size and was only re-post-trained, yet agent scores moved sharply, with Terminal Bench 2.1 rising from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4. It reportedly beats V4-Pro Preview on all nine published agent benchmarks while activating roughly a quarter of Pro's parameters per token, with the caveat that every number comes from DeepSeek's own harness.

DeepSeek moved V4 Flash out of preview on July 31, 2026 with the build name DeepSeek-V4-Flash-0731, keeping the same 284B-parameter MoE architecture and 13B active parameters per token, a 1M-token context window, and MIT-licensed weights seen in the April preview alongside the 1.6T-parameter V4 Pro. The changelog states the model was only re-post-trained, and the official API release is now in public beta. Three changes are attributed to the changelog: re-post-training rather than a new model, native Responses API support with the build specifically adapted for Codex so OpenAI Responses-format harnesses work without an adapter, and agent benchmark scores that pass V4-Pro-Preview. The update scope was the Flash API only, with the V4-Pro official release listed as coming soon.

DeepSeek's July 31, 2026 changelog carries two labels at once for DeepSeek-V4-Flash-0731: the checkpoint is the official release of V4 Flash, and the hosted API for it is in public beta. The April 24 Preview is superseded, and the stable API model ID deepseek-v4-flash did not change, so existing API integrations received the new checkpoint without any code change. The page explicitly declines to call the release generally available or production-stable. The same release day added a native Responses API aimed at Codex-style agents and official MIT-licensed open weights on Hugging Face. The service contract may still shift while in beta, since documented limits such as the default account concurrency of 2,500 for deepseek-v4-flash apply, and fields the Responses API currently ignores could later be honored or rejected.

DeepSeek published the official weights and model card for DeepSeek-V4-Flash-0731 on Hugging Face, describing it as the official release of DeepSeek-V4-Flash that supersedes the preview with substantially enhanced agentic capabilities. The model keeps the same structure as DeepSeek-V4-Flash-DSpark, including an attached speculative decoding module, and the card states it outperforms DeepSeek-V4-Pro (Preview) despite far fewer activated parameters. The model card lists vendor benchmark results including Terminal Bench 2.1 at 82.7 versus 61.8 for the preview and 72.1 for V4-Pro (Preview), plus DeepSWE at 54.4 and Toolathlon-Verified at 70.3. It notes that no Jinja chat template is included, that reasoning effort now supports low, high, and max levels, and that vLLM users enable DSpark speculative decoding via a --speculative-config flag with method dspark.