DeepSeek, the Hangzhou-based lab whose V3 release shipped at a fraction of frontier training costs, released DeepSeek-V4-Flash-0731 on Hugging Face on July 31. The card frames this as the official production version of V4-Flash, superseding the preview build, not an incremental update. Where an earlier report this month covered a preview version matching GPT-5.6 Luna at lower cost, this release comes straight from DeepSeek’s own documentation for a shipped model, not a leak or a third-party test.
The card lists nine agentic and coding benchmarks. On Terminal Bench 2.1, the new release scores 82.7, up from 61.8 for the Flash preview and ahead of the 72.1 posted by the Pro Preview model, DeepSeek’s larger sibling to Flash. On DeepSWE, a coding-agent benchmark, the gap widens: 54.4 for the new release against 12.8 for Pro Preview and 7.3 for the old Flash preview. Cybergym and Agents’ Last Exam follow the same pattern, with V4-Flash-0731 posting 76.7 and 25.2 against Pro Preview’s 52.7 and 16.5.
Against outside models the picture is mixed. V4-Flash-0731 edges past GLM-5.2, Zhipu AI’s model, on most of the nine benchmarks. Opus-4.8, from Anthropic’s Claude Opus line, still leads on most rows. On NL2Repo, Opus-4.8 scored 69.7 against Flash’s 54.2. On DSBench-Hard, Opus-4.8 scored 71.7 against Flash’s 59.6. DeepSeek calls the release “broadly competitive with the strongest proprietary models available.” The numbers back a qualified version of that claim, not an unqualified one.
The release also carries a speculative decoding module called DSpark, built into the model rather than left to whichever inference engine runs it. Speculative decoding works by having a small draft model propose several tokens at once, which the full model then checks in parallel instead of generating one token at a time. The output distribution does not change. Only the latency does, usually for the better. DeepSeek says the DSpark draft weights come from the same checkpoint as the main model, so vLLM and SGLang deployments turn it on with a single configuration flag rather than sourcing and maintaining a separate draft model. That is the notable engineering choice here: pushing a serving-level optimization into the weights release removes a common source of drift between the model a lab claims and the model a deployment actually runs.
The card’s central claim, that the smaller Flash model beats DeepSeek’s own larger Pro model, does not mean Flash is now the stronger system across the board. It shows that on this specific set of agentic and coding tests, run at DeepSeek’s maximum reasoning-effort setting, the Flash architecture and its training run beat a preview build of the Pro line while using considerably fewer active parameters. The nine benchmarks are all agentic or coding-agent tests, not a general capability suite, and the card does not state exact parameter counts for either model. These are DeepSeek’s own numbers, produced with DeepSeek’s own harness (described on the card as not yet released), and no independent benchmark or third-party reproduction is cited.
DeepSeek licenses the release under the MIT License, which permits commercial use without royalty. The card omits pricing entirely and gives no specific context-length figure, despite the accompanying technical report describing itself as an approach to million-token context. Deployment guidance is limited to vLLM and SGLang serving commands tested on multi-GPU nodes, plus a note that high and max reasoning-effort settings call for output lengths up to 384,000 tokens.
Teams evaluating agentic coding models should treat these benchmark gaps as a reason to run their own tests, not a substitute for running them. Anyone planning to build on DSpark should also track whether DeepSeek ships the promised harness and encoding documentation before locking a production pipeline to it.
DeepSeek published the DeepSeek-V4-Flash-0731 model card on Hugging Face on July 31, 2026.
