DeepSeek V4 Pro is a large Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active per pass, designed around advanced coding, tool use, and long-horizon agentic work. The official full release, identified as DeepSeek-V4-Pro-0813, builds on the earlier V4 Pro Preview model structure and adds a DSpark speculative decoding module, with the weights published openly under an MIT license and accompanied by a technical report on arXiv. That lineage matters in practice: the 0813 checkpoint is presented as a superseding release rather than a fresh design, and it focuses on sharpening agentic capabilities and production performance that were already present in the preview.

In benchmark comparisons shared by DeepSeek, the 0813 release improves on both the V4 Pro Preview and the V4 Flash variant across every listed test, including a jump on HLE with tools from the preview's 48.2 to 60.0, a Terminal Bench 2.1 score of 87.9, and an NL2Repo score of 61.5 that more than doubles the preview's 38.5. These gains position it as broadly competitive with strong proprietary peers such as Opus-4.8 and Kimi K3 on agentic coding and tool-use tasks. The release also exposes configurable reasoning effort (low, high, and max) and supports very long context windows in hosted deployments, making it a practical fit for teams that need an open-weight model capable of running large codebases, multi-step automation, and tool-driven research workflows end to end.