DeepSeek V4 Pro 0813 (Alibaba)

GMI Cloud's August 12, 2026 blog post confirms that DeepSeek V4 Pro left preview status on that date, with the official deepseek-v4-pro endpoint now pointing to the 0813 build as documented on DeepSeek's API and pricing pages. The article details the architecture as a 1.6 trillion-parameter Mixture-of-Experts model wit For developers and Eden AI integrators, the post reports that Pro pricing carries over from preview at approximately $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 to $0.93 per million output tokens depending on provider, with concurrency capped at 500 requests on Pro v

InferenceX's model page provides a deep architectural breakdown of the DeepSeek V4 series, sourced from the official DeepSeek-V4-Pro model card, the DeepSeek API news page dated April 24, 2026, and the arXiv technical report (arXiv:2606.19348). The model is described as a 61-layer MoE with 1.6T total and 49B active par For engineers consuming the qwen/deepseek-v4-pro-0813 endpoint via Eden AI, the practical implications are that the model combines FP4 experts with FP8 mixed precision, uses the Muon optimizer during training, applies YaRN RoPE to extend context to 1M tokens, has a vocabulary of 129,280 with embedding dimension 7,168,

This third-party aggregator blog consolidates the DeepSeek V4 Pro 0813 benchmark picture, listing the build dated 2026-08-12 with an Artificial Analysis Intelligence Index of 53 and Coding score of 69. It reports DeepSeek's vendor-published agentic numbers — Terminal-Bench 2.1 at 87.9, DeepSWE at 62.7, and CyberGym at For Eden AI users evaluating the qwen/deepseek-v4-pro-0813 endpoint, the relevant takeaways are that headline scores are vendor-reported under DeepSeek's new Harness framework and remain only partially replicated by third-party evaluators, and that broader model listings on the same page include nearby launches such as