Taishan-Vision

DeepSeek Harness text models cannot natively understand images, yet most users only need occasional image recognition without switching to a multimodal model or paying for extra vision API calls. Taishan Vision fills exactly that gap: it transparently invokes an external vision model to describe an image and feeds the result into the current model's context, so a pure-text model can still make sense of pictures.

By default the plugin connects to two completely free Zhipu GLM vision models (glm-4.6v-flash and glm-4.1v-thinking-flash), bringing recognition cost to zero. It also scans every DSH model that supports image input, so it is vendor-agnostic and works out of the box if you already have a vision model configured. Credential-aware routing keeps models without an API Key disabled by default to avoid repeated failed calls, while the Key can be pasted directly into a settings panel, persisted to a local credentials file without ever being echoed back in plaintext or lost on restart. The zhipu-glm provider declaration is fully built in and auto-written when missing, so a fresh install needs no manual configuration. A built-in taishan_diag tool and fenced output from the vision model further aid troubleshooting and guard against prompt-injection attacks.

Well suited for individual developers, researchers, and everyday users who work with text-only models (such as DeepSeek's text series) but occasionally need to read an image, especially when they want a zero-config, zero-cost, ready-to-use experience.