DeepSeek V4 Flash Vision Exp extends the Flash family with experimental multimodal vision understanding, accepting images alongside text so users can describe pictures, read text from screenshots, and analyze charts. It supports JPEG, PNG, GIF, and WebP inputs, with format detection taken from the actual file content rather than filenames or declared MIME types, and it slots into the existing OpenAI-compatible Chat Completions and Responses APIs against the standard DeepSeek endpoint. The vision guide documents three ways to feed images, including base64-encoded inline data within content-block arrays, so existing text-based integrations can be extended to visual inputs without changing their overall request shape.
According to the release notes, this experimental variant is positioned for agent-style and visual reasoning workloads, with reported scores such as Terminal Bench 2.1 at 83.9, DeepSWE at 59.3, DSBench-Hard at 63.6, ApexBench Pass@1 at 36.5, Chartography at 64.3, ZeroBench Pass@5 at 35.0, and Agents' Last Exam at 27.3, alongside NL2Repo at 57.7 and AutomationBench Public at 25.7. The release framing notes that pure-text capabilities remain on par with the base Flash model, while benchmarks that require visual understanding show a substantial lift, closing the gap to higher-tier reasoning systems on multimodal agent tasks. Practically, it is best suited to developers who want to add image-aware tool calling and structured outputs to agent pipelines without leaving the DeepSeek API surface.
