MindStudio published a benchmark comparison titled \"DeepSeek-V4-Pro-0813 Benchmarks: How It Stacks Up Against Opus and Kimi K3\" on August 16, 2026, which is the only on-page evidence available for this candidate. The supplied scrape is dominated by cookie-consent UI text, so no specific benchmark numbers, methodology Because the actual benchmark content was not retrievable in the supplied evidence, no quantitative claims about DeepSeek-V4-Pro-0813's performance against Claude Opus or Kimi K3 are asserted here. The candidate is accepted only as evidence that a third-party benchmark write-up exists for the 0813 build; substantive com

DeepSeek-V4-Pro-0813 was released by DeepSeek on August 13, 2026, thirteen days after the smaller DeepSeek-V4-Flash-0731, as recorded on the AI Release Tracker model page. The release is benchmarked across BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, CyberGym, and two addit The tracker page frames DeepSeek-V4-Pro-0813 as a major frontier-tier model available through multiple provider surfaces and ties it to benchmark coverage that emphasizes agentic coding and tool use. Because the Eden AI-routed databricks/databricks-deepseek-v4-pro-0813 catalog entry does not have a confirmed Eden-AI-sp

GMI Cloud reports on August 12, 2026 that the deepseek-v4-pro endpoint on DeepSeek's official API now points to the new DeepSeek-V4-Pro-0813 build, ending a preview window that opened with the model's April 24, 2026 debut. Specs carried over from the preview include 1.6T total parameters with about 49B active per token The blog highlights 0813's largest gains on agentic engineering benchmarks spanning multi-step coding, tool use, cybersecurity workflows, and full-stack development, evaluated using DeepSeek's upcoming Harness framework built around a minimal shell-and-editor environment. GMI Cloud lists pricing carried over from the p

DeepSeek V4 Pro is a 1.6-trillion-parameter Mixture-of-Experts model with roughly 49B active parameters per token and a native 1-million-token context window, documented in detail by the InferenceX architecture page. The 61-layer stack uses hybrid attention that alternates Compressed Sparse Attention and Heavily Compre According to InferenceX, DeepSeek released the V4 series on April 23, 2026 and open-sourced V4-Pro alongside the smaller V4-Flash under the MIT License, with dual Thinking/Non-Thinking modes and 1M context as defaults. The model is positioned as agentic and coding-first, described as open-source SOTA on agentic-coding