DeepSeek-V4-Flash-Vision-Exp Official ReleaseProvider report
2026-08-21Based on the complete comparison image on DeepSeek's official DeepSeek-V4-Flash-Vision-Exp release page dated August 21, 2026. The source contains three models and 11 benchmark rows, producing 31 available scores; dash cells do not create scores. For public code-agent text tasks, DeepSeek models use DeepSeek Harness in minimal mode with max effort, top_p=0.95, and temperature=1.0. DeepSeek-V4-Flash-0731 ignores multimodal elements in ApexBench and Agents' Last Exam, and those markers are retained in raw.