DeepSeek Launches V4-Flash-Vision-Exp on API, Beats Anthropic Opus 4.8 on Two Multimodal Benchmarks

DeepSeek released V4-Flash-Vision-Exp on its paid developer platform, adding image understanding to its 284B-parameter mixture-of-experts model that activates 13B per prompt. The experimental model surpassed the base V4-Flash on six of seven text benchmarks and scored over 10 points higher on two visual benchmarks, beating Anthropic's Opus 4.8 on ALE (1,000+ multi-step app tasks) and ZeroBench (100 hard image analysis tasks). It uses HCA/CSA compression that cut 1M-token processing cost by 73%.