ModelFarm Skill Installation and Usage

Introduction

The ModelFarm skill is a task-oriented skill guide for AI Agents. It helps Agents work with APLUX ModelFarm, including model lookup, model download, board deployment, demo execution, and troubleshooting.

This guide is not limited to a specific Agent implementation. Codex, Claude Code, custom Agents, or other tools that support skill, prompt, or workflow extensions can all use this guide to organize ModelFarm-related capabilities. Codex is used only as an example for the installation path and usage flow.

This skill applies to the following scenarios:

• Query whether a ModelFarm model supports a specific chipset, precision, or backend version
• Use mms on an APLUX board to log in, search, and download models
• Run model demos on QCS6490, QCS8550, QCS8625, IQ8, IQ9, and other platforms
• Run CV model demos with AidLite
• Run LLM / VLM local inference demos with AidGen
• Run OpenAI-compatible HTTP API services with AidGenSE
• Run ASR / TTS file demos with AidVoice

Installation

ModelFarm skill repository:

https://github.com/aplux-ai/model-farm-skill

Clone the repository locally, then copy it to the target Agent's skills directory.

git clone https://github.com/aplux-ai/model-farm-skill.git
mkdir -p <agent-skills-dir>
cp -r model-farm-skill <agent-skills-dir>/model-farm

For Codex:

git clone https://github.com/aplux-ai/model-farm-skill.git
mkdir -p ~/.codex/skills
cp -r model-farm-skill ~/.codex/skills/model-farm

Example directory structure after installing for Codex:

~/.codex/skills/
└── model-farm/
    ├── SKILL.md
    └── references/

Where:

• SKILL.md: the skill entry file that defines the trigger scope, workflow, and global rules
• references/: detailed instructions for different task types, such as model download, AidGen execution, and AidVoice execution

Verify Installation

After installation, confirm that SKILL.md exists.

Generic check:

ls <agent-skills-dir>/model-farm/SKILL.md

For Codex:

ls ~/.codex/skills/model-farm/SKILL.md

If the file is found, the skill has been placed in the correct location.

When starting a ModelFarm-related task in an Agent, mention keywords such as ModelFarm, model-farm skill, mms, AidGen, AidGenSE, or AidVoice. Agents that support automatic skill matching can load this skill based on the task content. Agents without automatic matching can import SKILL.md as a system prompt, tool description, or workflow guide.

Usage

After installation, directly describe the ModelFarm task you want the Agent to complete. The following are common request types.

1. Query Models

Query whether a model is available, or whether it supports a specific chipset, precision, or backend version.

Check whether Qwen3-4B on ModelFarm supports QCS8550.

Query Qwen2.5-VL-3B support for qnn2.36 on ModelFarm.

2. Download Models

Download specified model resources to a board or local directory.

Use the ModelFarm skill to download the QCS8550 qnn2.36 model package for Qwen3-4B.

Use the ModelFarm skill to download the QCS8550 w4a16 model package for Qwen2.5-VL-3B.

3. Run LLM / VLM Demos

Run board-side demos for large language models or vision-language models.

Use the ModelFarm skill to run the Qwen3-4B local chat demo on QCS8550.

Use the ModelFarm skill to run the Qwen2.5-VL-3B vision-language chat demo on the board.

4. Start an API Service

Start an OpenAI-compatible HTTP API service or Web UI.

Use the ModelFarm skill to start an OpenAI-compatible API service for Qwen3-4B.

Start the /v1/chat/completions service for Qwen3-4B on the board.

5. Run Voice Demos

Run ASR speech recognition or TTS speech synthesis demos.

Use the ModelFarm skill to run the SenseVoiceSmall ASR file recognition demo on the board.

Use the ModelFarm skill to test the MeloTTS text-to-speech demo.

Board Connection Requirements

Board-side execution requires the device IP address, username, and password.

If only the IP address is provided, the Rhino-Pi default account is usually used:

Username: aidlux
Password: aidlux

Common working directory variables:

BOARD_WORKDIR="${BOARD_WORKDIR:-$HOME}"
AIDLLM_DIR="$BOARD_WORKDIR/aidllm"
AIDMLM_DIR="$BOARD_WORKDIR/aidmlm"
AIDVOICE_DIR="$BOARD_WORKDIR/aidvoice"

Notes

• ModelFarm model directories, package names, chipset support, and performance data may change. Prefer real-time query results.
• Do not assume support for a chipset, precision, or QNN version before confirming it.
• After downloading a model, inspect the extracted README, scripts, and configuration files before running commands.
• LLM / VLM local deployment defaults to AidGen.
• Use AidGenSE for OpenAI-compatible HTTP service requests.
• CV model demos usually use AidLite.
• ASR / TTS file demos usually use AidVoice.
• Monitor progress during long-running commands such as mms get and aidllm pull api.