DeepSeek V4 Pro 0813 (TensorX)

AI Release Tracker's entry confirms DeepSeek-V4-Pro-0813 as a DeepSeek model released on Thursday, August 13, 2026, thirteen days after DeepSeek-V4-Flash-0731, with benchmark coverage spanning BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, and CyberGym plus two additional sui The tracker also explicitly notes that thinking and non-thinking modes share the same published price, so developers do not need to switch rate cards when toggling reasoning effort. Pricing is presented as sourced from DeepSeek's own API documentation, making the page a useful aggregator for the official peak/off-peak

AIToolsReview's feature frames DeepSeek-V4-Pro-0813 as a quieter production build focused on serving efficiency rather than a ground-up redesign, adding roughly 51.7 billion additional parameters across four new DSpark speculative-decoding keys on top of the April preview. It reports DeepSeek's own Terminal-Bench 2.1 c The review underscores a benchmark reproducibility gap that materially affects how developers should weight DeepSeek's published agent and coding scores, and it carries through the off-peak/peak pricing tiers of $0.66/$1.32 cache-miss input and $1.98/$3.96 output per million tokens. Together with the MIT-licensed open

DataLearner.com's model page describes DeepSeek-V4-Pro-0813 as a text-only 1.6T-total/49B-active MoE with a 1M-token context window and 384K-token maximum output, released on August 13, 2026 with a May 2025 knowledge cutoff. The page distinguishes Normal, Think High, and Think Max reasoning modes, lists responses-API a For benchmarking, the page reports LiveCodeBench at 93.50 (ranked 1st of 127), MMLU Pro at 87.50 (ranked 11th of 133), and SWE-bench Verified at 80.60 (ranked 11th of 115). It also flags an important caveat for anyone self-hosting: the public Hugging Face checkpoint remains labeled as the April preview and is not repre

OpenLLMStack's profile documents DeepSeek-V4-Pro-0813 as the official release of DeepSeek's 1.6-trillion-parameter Mixture-of-Experts flagship, with 49B active parameters, a 1M-token context window, FP4+FP8 mixed precision, and a DeepseekV4ForCausalLM architecture using Compressed Sparse plus Heavily Compressed Attenti For developer integration the page highlights a selectable reasoning-effort parameter (low, high, or max), with DeepSeek recommending temperature 1.0 and top-p 0.95 for agentic work and a 384K-token maximum output at higher settings. It also confirms native OpenAI Responses API support with a one-click Codex setup, mat