Introduction

DeepSeek Harness (DSH) is a plugin-centric framework. When locally deploying Qwen models, developers usually need to manually handle vLLM configuration and integration with the DSH ecosystem. The dsh-llm-qwen-local plugin provides a standard LLM adapter layer specifically for driving locally deployed Qwen models. It exposes an OpenAI-compatible interface through a vLLM backend, allowing existing DSH workflows to call local models directly, while providing a visual Web configuration interface.

Core Features

  • Local deployment adaptation: Provides a standard adapter layer for local Qwen models.
  • vLLM backend support: Relies on a vLLM service for inference.
  • OpenAI-compatible API: Exposes the standard /v1/chat/completions endpoint.
  • Multimodal toggle: Supports configuring the multimodal switch on a per-model basis.
  • Reasoning effort configuration: Fully configurable reasoning effort levels, including the default effort and how the off state is expressed.
  • Web settings page: Provides a bilingual (Chinese/English) Web settings page with model discovery and configuration.
  • Image projection: Supports requesting image projection.
  • Zero runtime dependencies: Removes dependency on @deepseek-ai ecosystem runtime packages (only eventsource-parser and Node.js native modules are retained).

Installation and Enablement

Before installing, ensure that the following system requirements are met:
* DSH version: 0.1.2-rc.1 or later.
* Node.js version: 18 or later.
* vLLM instance: started and running an OpenAI-compatible endpoint.
* Profile: use the web or headless profile (they include @deepseek-ai/dsh-attachment by default).

Install command:

dsh plugin --profile web add dsh-llm-qwen-local

Note: DSH 0.1.1-rc.2 and earlier will cause startup failure; the Web application will report a startup error indicating that it is waiting for services such as remote.credentials. Upgrade DSH or remove the plugin.

Typical Usage

After installation, define the adapter in the DSH configuration file. The main configuration items include baseURL (pointing to the vLLM endpoint), the models list, and the reasoning configuration.

- id: llm-qwen-local
  name: dsh-llm-qwen-local
  config:
    baseURL: http://127.0.0.1:8000/v1
    models:
      - id: qwen3.8
        name: Qwen3.8 (local)
        multimodal: true
        reasoning:
          efforts:
            - { id: off, wire: none }
            - { id: low, wire: low }
            - { id: medium, wire: medium }
            - { id: xhigh, wire: xhigh }
          defaultEffort: xhigh

In the configuration above:
1. baseURL points to the OpenAI-compatible interface of vLLM.
2. The multimodal switch controls whether the current model supports image input.
3. The reasoning section defines the mapping of reasoning effort levels; the wire field corresponds to the parameter name accepted by vLLM, and defaultEffort sets the default effort level.

Notes

  • Version compatibility: The plugin strictly depends on DSH 0.1.2-rc.1 or later. Older DSH versions do not support the remote-namespace client model required by the plugin.
  • vLLM startup parameters: To correctly parse reasoning blocks and tool calls, the vLLM instance must be started with the following specific parameters:
    --reasoning-parser qwen3
    --enable-auto-tool-choice --tool-call-parser qwen3_coder
    --max-model-len 262144
  • License and maintenance: The plugin is licensed under MIT and maintained by the community. It is not an official product of DeepSeek or Qwen.
  • Configuration cleanup: The legacy maxRequestImageBytes configuration item has been removed. Any leftover occurrences of this field in existing configuration files will be ignored, and no manual migration is required.

Summary

dsh-llm-qwen-local provides developers with a lightweight and fully configurable solution for running Qwen models locally and integrating them with the DSH ecosystem. It simplifies debugging of reasoning effort and multimodal configuration through the Web interface, while reducing runtime complexity through a zero-dependency design. For more configuration details and design documentation, refer to the project repository.

Plugin Directory Page
GitHub Repository