dsh-official-vision

Vision inference in DeepSeek Harness has long relied on third-party bridges — GLM, Gemini, InternVL, and others. They work, but every image request detours through an external model, adding latency and muddying the billing and privacy boundary. dsh-official-vision solves that in one move: it registers the official multimodal model deepseek-v4-flash-vision-exp as a native provider route so that encoded images are sent straight to api.deepseek.com with no third-party model in the path.

Under the hood the plugin supports two image-transport modes. The default inline base64 path is ideal for local images up to 32 MiB per file. When preferFilesApi is enabled, images are first uploaded via the official Files API, cached by a digest-based file_id for deduplication, and then referenced in subsequent requests — a cleaner fit for large files or the same image reused across turns. Plain-text conversations are transparently forwarded to the standard deepseek route, so there is zero overhead when no image is attached.

It is best suited for developers running a Harness build that has not yet shipped a built-in vision route and want to skip the third-party detour entirely, as well as teams that need a single, auditable, official entry point for rate-limiting, logging, and cost tracking. Once installed, you simply pick DeepSeek-V4-Flash-Vision-Exp from the model selector, paste an image, and the entire inference chain stays on DeepSeek's own infrastructure.