AI BriefingDeepSeek-V4-Pro-0813: DeepSeek-V4-Pro with DSpark: 87.9 on Terminal Bench
deepseek-ai/DeepSeek-V4-Pro-0813
About the project
DeepSeek-V4-Pro-0813 is the official release model that replaces the previous preview version. It features significantly improved agent capabilities and overall performance for production environments. In particular, the integration of the DSpark speculative decoding module enhances inference speed and efficiency.
It achieved a score of 87.9 on Terminal Bench 2.1 and 61.5 on NL2Repo, demonstrating strong performance in code agent tasks. On the HLE benchmark, it reached 60.0 with tool use. Its performance remains competitive with the latest rival models such as GLM-5.2, Kimi K3, and Opus-4.8.
Instead of Jinja-based chat templates, it provides dedicated Python scripts for message encoding and parsing in an OpenAI-compatible format. The reasoning_effort parameter can be adjusted across three levels—low, high, and max—to control how deeply the model thinks before answering. In vLLM environments, DSpark inference can be enabled with a single --speculative-config flag.
Released under the MIT license, it is free to use in commercial projects. It supports 8-bit and fp8 precision options and is immediately available from major inference providers including Together, Novita, Baseten, and Fireworks AI. It is suitable for developers who need complex agent workflows or large-scale codebase analysis.
deepseek-ai/DeepSeek-V4-Pro-0813
The original page has no description.
text-generation
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.