DeepSeek Harness (DSH) has a plugin architecture that allows developers to extend model support. In daily development, integrating external APIs requires additional configuration. This plugin specifically targets the free model tier of AMD Radeon Token Factory, resolving the configuration issue of directly calling these GPU Cloud models in DSH.

Plugin Overview

This is a DeepSeek Harness plugin that probes the free models on AMD Radeon Token Factory daily. It is maintained by jiuge613 and released under the MIT license. It focuses on integrating free-tier models and does not include paid-tier content. The plugin is purely drop-in, requires no build step, and has zero dependencies.

Core Features

  • Daily Probing: Automatically detects which models are available in the free tier of AMD Radeon Token Factory.
  • Supported Models: Supports DeepSeek V4, MiMo V2.6, Qwen3.8, and GLM 5.3.
  • Local Forwarding: Provides an OpenAI-compatible local forwarding port and supports calls via base URL + Key.
  • Real Effect: Thinking intensity (Light / Balanced / Deep) truly corresponds to the actual max_tokens budget issued.
  • Load Monitoring: Exposes load monitoring and displays idle/busy/full status along with utilization.
  • In-App Upgrade: Supports one-click upgrades and hot reloading.
  • Usage Dashboard: Locally records Token heatmaps, total volume curves, and output speed.
  • Trust Fence: Plugin routing takes priority over kernel dispatch and includes built-in access checks.

Installation and Enablement

Desktop

  1. Open DSH left navigation bar → Plugins.
  2. Click the Add Plugin button in the upper-right corner.
  3. Paste the repository address into the input box:
    https://github.com/jiuge613/dsh-amd-free-model
  1. Confirm installation and restart the app.

Command Line

Run the installation command with the corresponding profile:

dsh plugin --profile web add github:jiuge613/dsh-amd-free-model

Restart the app after installation is complete.

Typical Usage

  1. Fill in an AMD Radeon Token Factory API Key on the DeepSeek Harness settings page.
  2. The plugin probes the AMD free tier once daily to detect available models.
  3. Call models through the local forwarding port using base URL + Key.
  4. Upgrade the plugin with one click on the settings page.

Notes

  • Model Scope: Only free models are listed; paid-tier models are not included.
  • 429 Errors: A 429 error indicates that compute capacity is busy, not that your quota has been exhausted. There is no need to change the Key; just retry later.
  • Key Security: The Key is saved only locally with 0600 file permissions and never appears in responses.
  • Loopback Listening: The forwarding listener binds to loopback only; non-loopback addresses are rejected directly.
  • Data Privacy: The usage dashboard remains entirely local, and no data is uploaded.

Summary

This plugin uses daily probing and local forwarding to make it easy for DSH users to use AMD’s free GPU models. The plugin maintains a detailed model list and load status, and all sensitive data is stored locally. Source code updates and more information can be found in its GitHub repository.