Skip to content

deepseek-harness-SupportVisionModel

⭐ 8 · ✅ active · plugin
Type plugin Category Vision & multimodal
Stars ⭐ 8 Status ✅ active
Author TryDing-T Updated 2026-08-14
Subcategory 👁️ Vision tools Capabilities multimodal, vision

One-liner

Secondary development of deepseek-harness supporting a separately configured vision model for reading images.

About

DeepSeek Harness SupportVisionModel is a community fork of DeepSeek Harness. It keeps the upstream plugin architecture and adds a unified image-routing path for image-capable and text-only main models. This repository is maintained independently at TryDing-T/deepseek-harness-SupportVisionModel. It is not an official DeepSeek AI distribution.

✨ Key Features

  • Model catalog entries can explicitly declare text-and-image input support.
  • A dedicated Vision settings page selects the auxiliary provider and model, maximum attempts, per-attempt timeout, and output-token limit.
  • Image-capable main models receive the original image attachments.
  • Text-only or unknown-capability main models receive a validated description from the configured vision model.
  • Missing or exhausted vision routes degrade to explicit attachment references instead of rejecting a successful user or tool action.
  • Web uploads, top-level tool image results, ordinary read_image, and nested Code Mode read_image share the same router.

📦 Install

git clone https://github.com/TryDing-T/deepseek-harness-SupportVisionModel.git
cd deepseek-harness-SupportVisionModel
pnpm install
pnpm run build
pnpm dsh web

🚀 Quick Start

git fetch upstream
git diff main..upstream/master

📚 Learn more

Configure the vision route

The auxiliary model must explicitly declare both text and image input support. One attempt sends the complete ordered image batch without tools.

Configure image routing

  1. Open Settings → Models, configure the provider, model, and API key, then enable Supports images only for a real multimodal endpoint. 2. Open Settings → Vision, select that provider and model, and set the attempt, timeout, and output-token limits. 3. Select any main model. Image-capable models receive images directly; text-only models use the configured vision route. API keys belong