DeepSeek V4 Pro is an open-weights large language model from DeepSeek-AI, released under an MIT license and publicly hosted on the DeepSeek-AI Hugging Face organization, with an NVIDIA NGC NIM container available for enterprise deployment. The newer 0813 checkpoint is presented as the official release that supersedes an earlier preview, reusing the preview model structure and attaching a DSpark speculative decoding module to accelerate inference. A linked technical report on arXiv (identifier 2606.19348) accompanies the release and documents the architecture and evaluation methodology behind the family.

The 0813 release emphasizes enhanced agentic capability and production-grade performance, with the model card reporting benchmark results that show it outperforming the V4-Pro preview on tasks such as HLE (with and without tools), Terminal-Bench 2.1, and NL2Repo. These gains position it as broadly competitive with strong proprietary systems on tool-assisted reasoning and code-oriented agent benchmarks, making it a practical fit for coding assistants, tool-using agents, and other production workflows where speculative decoding can meaningfully improve throughput.