{"id":4636,"date":"2026-09-26T18:16:40","date_gmt":"2026-09-26T09:16:40","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/model-deepseek-ai-deepseek-v4-en\/"},"modified":"2026-09-28T05:32:19","modified_gmt":"2026-09-27T20:32:19","slug":"model-deepseek-ai-deepseek-v4-en","status":"publish","type":"page","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-deepseek-ai-deepseek-v4-en\/","title":{"rendered":"DeepSeek-V4 Guide: VRAM Requirements, GGUF Builds"},"content":{"rendered":"

About This Model<\/h2>\n

DeepSeek-V4<\/strong> is DeepSeek’s MoE model series, built to be “highly efficient” enough to handle million-token contexts routinely. The centerpiece of this page, DeepSeek-V4-Pro-0813<\/strong>, is the official release of the flagship DeepSeek-V4-Pro and replaces the earlier preview. It keeps the preview’s architecture and adds a DSpark<\/strong> module for speculative decoding.<\/p>\n

According to the technical report’s abstract, DeepSeek-V4-Pro is an MoE model with 1.6T<\/strong> total parameters and 49B<\/strong> active per token, pre-trained on more than 32T tokens. It combines two compressed attention mechanisms (CSA and HCA) for long-context efficiency. At a one-million-token context, it needs only 27% of the per-token inference FLOPs and 10% of the KV cache of the previous-generation DeepSeek-V3.2. Reasoning depth can be set to low<\/code>, high<\/code> or max<\/code> with reasoning_effort<\/code>.<\/p>\n

What Makes It Stand Out<\/h2>\n

1. Agentic performance improved sharply over the preview.<\/strong> An excerpt from the publisher’s comparison table (bold marks the best value in each row):<\/p>\n

\n

\u2192 Scroll horizontally to see all columns<\/p>\n\n\n\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nV4-Pro-0813<\/th>\nV4-Pro (Preview)<\/th>\nGLM-5.2<\/th>\nKimi K3<\/th>\nOpus-4.8<\/th>\nFable-5 (w\/ fallback)<\/th>\n<\/tr>\n<\/thead>\n
Terminal Bench 2.1<\/td>\n87.9<\/td>\n72.1<\/td>\n81.0<\/td>\n88.3<\/strong><\/td>\n85.0<\/td>\n88.0<\/td>\n<\/tr>\n
DeepSWE<\/td>\n62.7<\/td>\n12.8<\/td>\n46.2<\/td>\n67.5<\/td>\n58.0<\/td>\n70.0<\/strong><\/td>\n<\/tr>\n
Cybergym<\/td>\n83.3<\/strong><\/td>\n52.7<\/td>\n\u2014<\/td>\n80.0<\/td>\n78.3<\/td>\n83.1<\/td>\n<\/tr>\n
AutomationBench (Public)<\/td>\n31.8<\/strong><\/td>\n12.8<\/td>\n12.9<\/td>\n30.8<\/td>\n27.2<\/td>\n29.1<\/td>\n<\/tr>\n
Toolathlon-Verified<\/td>\n74.1<\/td>\n55.9<\/td>\n59.9<\/td>\n76.5<\/td>\n76.2<\/td>\n77.9<\/strong><\/td>\n<\/tr>\n
NL2Repo<\/td>\n61.5<\/td>\n38.5<\/td>\n48.9<\/td>\n\u2014<\/td>\n69.7<\/strong><\/td>\n\u2014<\/td>\n<\/tr>\n
HLE (without \/ with tools)<\/td>\n42.7 \/ 60.0<\/td>\n37.7 \/ 48.2<\/td>\n40.5 \/ 54.7<\/td>\n43.5 \/ 56.0<\/td>\n49.8 \/ 57.9<\/td>\n53.3 \/ 63.0<\/strong><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n