Introduction

In the development of DeepSeek Harness (DSH), the default pure-text model cannot directly process image inputs. The Taishan Vision plugin calls Zhipu GLM’s free vision models to inject image recognition capability into the current model. It does not require the main model to support image inputs and directly solves the problem that pure-text models cannot “see” images.

Plugin Introduction

Taishan Vision is a DeepSeek Harness plugin maintained by iguanren. Its core function is to use Zhipu GLM’s free vision models (such as glm-4.6v-flash and glm-4.1v-thinking-flash) to inject image recognition capability into DSH’s pure-text models, enabling them to process image content.

Core Features

  • Free image recognition: Supports Zhipu GLM’s free vision models, with good Chinese recognition quality and automatic fallback support.
  • Native card panel: The card appears in the “Settings → Plugins → Plugin Configuration” area of DSH, matching the style of built-in cards and supporting collapse.
  • Ease of use: Supports one-click enable/disable, direct API Key configuration (without echoing the plaintext), model selection, and drag-and-drop reordering.
  • Security: The key is stored only in the credentials file and does not appear in logs or panels.
  • Tools and maintenance: Provides the taishan_diag diagnostic tool, and supports manually checking for updates and rescanning.

Installation and Enablement

Before installation, ensure that pnpm is available in the PATH (you can install it with npm i -g pnpm) and that Git is installed.

dsh plugin --profile web add github:iguanren/taishan-vision#v5.0

After installation is complete, restart the DSH interface and force refresh the browser (Ctrl+Shift+R). The card will appear under “Settings → Plugins → Plugin Configuration”.

Typical Usage

  1. Send a message containing an image in the conversation.
  2. The plugin automatically injects an image recognition instruction. The current model (a pure-text model) calls the taishan_describe_image tool, passing the image attachment ID to obtain a description.
  3. The model generates the reply based on the description. You can also directly ask the model in the prompt to “describe this image”.

Use Cases and Notes

It is suitable for scenarios that require adding image understanding capabilities to pure-text models such as DeepSeek without replacing the main model.

Notes:
* Before use, configure the Zhipu API Key under “Plugin Configuration”.
* The plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license before installation.
* If the card is not displayed, try restarting DSH and force refreshing the browser.

Conclusion

Taishan Vision integrates Zhipu GLM’s free vision models to add image processing capabilities to the pure-text models of DeepSeek Harness at low cost. The plugin panel is well integrated and feature complete, making it suitable for use in local private deployment environments.

Project URL: https://github.com/iguanren/Taishan-Vision