DeepSeek's vision model enters the agent race

DeepSeek released the experimental V4-Flash-Vision-Exp on Aug. 21, 2026, making it available through its API with image input and agent-tool support. The model keeps the reported text capabilities of V4-Flash, while DeepSeek's own table puts it close to Opus-4.8 on multimodal agent benchmarks. 1

The useful read is narrower than the launch headline: the model wins three of 11 published comparisons, trails on eight, and has a 12-point gap on NL2Repo. The benchmark footnote also says text-only V4-Flash ignores multimodal elements in two of the visual tests. 23