DeepSeek has published the official release of DeepSeek-V4-Pro, a versioned checkpoint it calls DeepSeek-V4-Pro-0813, replacing the preview the company shipped earlier this year. The model is available now under an MIT license with weights hosted on Hugging Face, where the official model card describes the release as a production-focused update with a new speculative decoding module and sharp gains on agentic benchmarks. Reuters confirmed the release on August 13, 2026, reporting that DeepSeek is stepping up expansion as demand for its models surges.
The 0813 checkpoint keeps the DeepSeek-V4-Pro (Preview) architecture and adds a DSpark speculative decoding module, which lets the model draft and verify tokens in parallel to speed up inference. DeepSeek positions the release around agentic workloads: the model card says agentic capabilities are “greatly enhanced” and that performance improvements are “especially pronounced in production environments.” The reasoning effort parameter now supports three levels (low, high, and max), and DeepSeek recommends a maximum output length of 384K tokens when running at the high or max settings.
Benchmarks against the preview and the Flash variant
The model card includes a benchmark table that compares DeepSeek-V4-Pro-0813 against the V4-Pro preview, the V4-Flash-0731 release, and several proprietary models. On Humanity’s Last Exam, the new checkpoint scores 42.7 without tools and 60.0 with tools, up from 37.7 and 48.2 on the preview. On Terminal Bench 2.1 it moves to 87.9 from 72.1, and on DeepSWE, a software engineering benchmark, it jumps from 12.8 to 62.7. The largest relative gains cluster in agentic and coding tasks: NL2Repo rises from 38.5 to 61.5, Cybergym from 52.7 to 83.3, and AutomationBench from 12.8 to 31.8. DeepSeek says the new model is broadly competitive with the strongest proprietary models available, while noting that its code-agent results were produced with its own DeepSeek Harness framework at maximum reasoning effort.
The update also flows into the API. DeepSeek’s documentation now maps the deepseek-v4-pro model name to DeepSeek-V4-Pro-0813, and the deepseek-v4-flash name to an updated DeepSeek-V4-Flash-0731, with no changes to the calling convention. That matters for teams already on the platform, because it means an API rollout of the official release can happen without code changes. It also follows a period of pricing pressure: DeepSeek signaled a major API price hike in early August amid surging demand, which we covered in our report on the pricing change.
For developers who want to run the weights themselves, DeepSeek documents both vLLM and SGLang paths. DSpark is enabled in vLLM with a single speculative configuration flag using seven speculative tokens, and the reference command serves the model on a single four-GPU GB300 node. In SGLang the same mechanism is enabled with the --speculative-algorithm DSPARK flag, with the draft weights coming from the same checkpoint rather than a separate draft model.
The release carries a technical report on arXiv (2606.19348) tied to the broader DeepSeek-V4 line, whose stated goal is “highly efficient million-token context intelligence.” The 0813 checkpoint is the first official, versioned release of the Pro tier, and it arrives alongside the smaller Flash variant we analyzed when DeepSeek-V4-Flash-0731 shipped with its speed-over-size profile.
With the official weights, the MIT license, and the speculative decoding path now documented, DeepSeek has effectively closed the loop that began with the V4 preview: an open release, a hosted API, and a reference deployment stack that fits on a single server node. The next question is pricing, which DeepSeek has already signaled it intends to revisit as demand keeps climbing.
