AI Agent Hub
Back to plugins
🤖

dsh-llm-deepseek-vision

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install NagasakiSoyo-ui/dsh-llm-deepseek-vision

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install NagasakiSoyo-ui/dsh-llm-deepseek-vision in your terminal to install the vision plugin; the full source is available at https://github.com/NagasakiSoyo-ui/dsh-llm-deepseek-vision

About this plugin

DeepSeek v4-flash and v4-pro are powerful pure-text reasoning models, but they simply cannot process images. The dsh-llm-deepseek-vision plugin gives them a pair of eyes, enabling image understanding without touching the text reasoning core.

The mechanism is straightforward: the plugin registers a DeepSeek (Vision) provider route inside Harness. When a request carries no images, traffic passes through to the default text-inference route (opencode-go) with zero overhead. When images are attached, a dedicated vision model (mimo-v2.5 by default) first converts each image into a textual description, and the text model then reasons over those descriptions. The text inference engine never touches raw image bytes, keeping responsibilities clean and decoupled.

Ideal for developers working in the DeepSeek Harness environment who run pure-text models but need multimodal conversation support. The plugin exposes configuration for the vision provider, vision model, description prompt, per-call output token cap, and the delegate inference route, so you can swap in a cheaper or more nuanced vision model to match your cost and accuracy requirements.

Use Cases

  • Attaching images to DeepSeek text-model conversations for Q&A
  • Understanding user-uploaded screenshots, charts, or photos in context
  • Extending a text-only inference stack with image input without modifying the core

Best For

  • Developers deploying DeepSeek text-model services via Harness
  • Product teams needing image-text mixed chat on top of text-only models
  • Engineers who want to swap vision models flexibly to optimize inference cost