Introduction

The DeepSeek Harness (DSH) plugin ecosystem allows developers to extend capabilities. Most visual plugins currently available on the market achieve multimodal interaction by bridging third-party models (such as GLM, Gemini, and InternVL). This plugin fills that gap: it directly integrates with the official DeepSeek vision API, registers the deepseek-v4-flash-vision-exp route, and streams images to api.deepseek.com without passing through any intermediate model.

What Is This

  • Name: dsh-official-vision
  • Maintainer: wilianyichen
  • Category: Model inference
  • Positioning: A DSH plugin that directly connects to the official vision API and registers the official multimodal model as a provider route.

Core Features

  • Direct official API access: Images are sent directly to the official DeepSeek interface, with no third-party model involved.
  • Model registration: Automatically registers the deepseek-v4-flash-vision-exp route.
  • Format support: Supports JPEG, PNG, GIF, and WebP formats.
  • Transmission methods:
    • Base64 inline (default): Suitable for local images, with a single image ≤ 32MiB.
    • Files API: Suitable for large images (> 32MiB), with single images up to 64MiB.
  • No-image passthrough: For pure-text requests, it automatically passes through to the official deepseek route with no additional overhead.

Installation and Enablement

Use the official install command to add the plugin:

dsh plugin --profile web add dsh-official-vision

Configuration

Configure the following items in the DSH settings interface (Plugins -> dsh-official-vision) or in the profile’s cordis.patch.yml:

- id: dsh-official-vision
  name: 'dsh-official-vision'
  config:
    apiKeyEnv: DEEPSEEK_API_KEY      # 环境变量名，需提前配置
    baseURL: https://api.deepseek.com
    delegateProvider: deepseek       # 无图请求委托的路由
    detail: auto                     # low | high | original | auto
    maxImagesPerRequest: 600         # 单次请求最大图片数
    preferFilesApi: false            # true = 使用 Files API 处理大图

Typical Usage

  1. Select the DeepSeek-V4-Flash-Vision-Exp route in the model selector.
  2. Paste images directly into the input box to ask questions.

Notes and Limitations

  • Billing: Images are billed by tokens, with each image converted to at most 384 tokens, at the same price as V4-Flash.
  • Size limits:
    • Request body ≤ 48MiB
    • Base64 mode: single image ≤ 32MiB
    • Files API mode: single image ≤ 64MiB
    • Maximum images per request ≤ 600
  • Dependency requirements:
    • @deepseek-ai/cordis >= 4.0.1
    • @deepseek-ai/dsh-invariants >= 0.1.0-rc.5
    • @deepseek-ai/dsh-llm >= 0.1.0-rc.5

Conclusion

This plugin addresses the issue that existing visual plugins depend on third-party models, providing a direct path to DeepSeek’s official multimodal capabilities. The source code and directory information are as follows:
* Directory page: https://www.skillhub.cn/plugins/wilianyichen/dsh-official-vision
* GitHub: https://github.com/wilianyichen/dsh-official-vision