DeepSeek has launched DeepSeek-V4-Pro-0813, the official release of DeepSeek-V4-Pro that supersedes the preview version. The model is built on the DeepSeek-V4-Pro Preview architecture with a DSpark speculative decoding module and includes improvements for agentic and production workloads. DeepSeek says V4-Pro-0813 performs better than the preview version across the listed benchmarks.
DeepSeek-V4-Pro: Agent and coding performance
DeepSeek-V4-Pro-0813 shows higher scores than the preview model across the public and internal benchmarks provided by DeepSeek. It scores 87.9 on Terminal Bench 2.1, compared with 72.1 for V4-Pro Preview, while its scores on NL2Repo, Cybergym and DeepSWE are 61.5, 83.3 and 62.7, respectively.
The model scores 74.1 on Toolathlon-Verified, 25.7 on Agents’ Last Exam and 31.8 on AutomationBench (Public). On DeepSeek’s internal tests, it scores 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard, compared with 41.8 and 31.1 for the preview model.
For the code-agent tasks in the public benchmarks, DeepSeek evaluated V4-Pro-0813 using the minimal mode of DeepSeek Harness as the agent framework, with max reasoning effort, temperature set to 1.0 and top-p set to 0.95. DeepSeek identifies DSBench-FullStack as an internal full-stack development test set and DSBench-Hard as an internal test set of difficult coding-agent problems.
Flexible reasoning and API support
The reasoning_effort parameter now supports three levels:
- Low: Intended for simple tasks.
- High: Intended for daily agent workflows.
- Max: Intended for complex tasks requiring more deliberation.
The V4 lineup also adds native OpenAI Responses API support and is optimized for Codex, with a one-click setup. The API model names remain unchanged, with DeepSeek directing developers to its API documentation for setup details.
The release does not include a Jinja-format chat template. Instead, DeepSeek provides a dedicated encoding folder containing Python scripts and test cases for converting OpenAI-compatible messages into input strings and parsing the model’s text output. The encoding workflow also allows developers to specify the reasoning effort used for a request.
DSpark speculative decoding
DeepSeek-V4-Pro-0813 includes DSpark speculative decoding. Developers can use it with:
- vLLM: DSpark can be enabled through the
--speculative-configoption. DeepSeek’s example uses seven speculative tokens with greedy draft sampling. - SGLang: The
DSPARKspeculative algorithm can be enabled without specifying a separate draft model path, as the target and draft weights come from the same checkpoint. - Local deployment: DeepSeek provides an inference folder with instructions for model weight conversion and interactive chat demos.
Local deployment settings
DeepSeek recommends the following sampling settings for local deployment:
- Agentic scenarios: Temperature 1.0 and top-p 0.95.
- Other scenarios: Temperature 1.0 and top-p 1.0.
- High and max reasoning: Maximum output length of 384K tokens.
API pricing
DeepSeek is introducing separate peak and off-peak API rates with the V4 lineup. Off-peak pricing is 50% lower than peak pricing.
For DeepSeek-V4-Pro, the new rates are:
| Pricing | Peak | Off-peak |
| Cached input | 0.30 yuan / 1M tokens | 0.15 yuan / 1M tokens |
| Uncached input | 9 yuan / 1M tokens | 4.5 yuan / 1M tokens |
| Output | 27 yuan / 1M tokens | 13.5 yuan / 1M tokens |
The new pricing takes effect at 16:00 UTC on August 16, 2026.
Availability
DeepSeek-V4-Pro-0813 is available on the DeepSeek app and web through Expert Mode. It is also available through the DeepSeek API, with the existing API model names retained.
DeepSeek has provided API documentation for the setup process. The repository and model weights are released under the MIT License.