This project demonstrates how to use the qwen-agent library to interact with a Qwen 3 language model, potentially leveraging MLX for local execution and incorporating tools like web search.

Description

The core example (agentic_search.py) sets up an AI assistant powered by a specified Qwen 3 model (e.g., qwen3:0.6B running locally via Ollama). It showcases how to:

Configure the connection to the language model (local or API-based).
Define and integrate tools (like code_interpreter and MCP-based tools like DuckDuckGo search).
Run the agent with a user query and stream the responses.
Example prompt included - it was difficult to get the model with agents to also output the hyperlinks.

The smaller try6b.py is meant to test that qwen models work locally. 0.6B is a very small model, so easy to download, and was actually a a lot of fun to use. Event working with tools, when used with agentic_search.py worked, up to a point.

Prerequisites

Python 3.13 or higher
uv package manager installed.
A running Qwen 3 model endpoint (like one served by Ollama at http://localhost:11434) or appropriate API keys if using a hosted service.
Optional: mise for managing development tools (like Node.js, though its specific use in this Python project isn't detailed in the provided context).

Installation

Clone the repository:

git clone <your-repository-url>
cd qwen3

Set up a virtual environment (recommended):

# Using Python's built-in venv
python -m venv .venv 
source .venv/bin/activate # On Windows use `.venv\Scripts\activate`

# Or using uv to create and activate
uv venv .venv
source .venv/bin/activate # On Windows use `.venv\Scripts\activate`

Install dependencies using uv: The project dependencies are listed in pyproject.toml. Install the project and its dependencies:
```
uv pip install .
```
Alternatively, install dependencies directly from the list:
```
uv pip install mcp>=1.6.0 mlx>=0.25.1 mlx-lm>=0.24.0 python-dateutil>=2.9.0.post0 python-dotenv>=1.1.0 "qwen-agent[code-interpreter]>=0.0.20"
```

Usage

Configure the LLM:
- Modify agentic_search.py to point to your Qwen 3 model endpoint or provide necessary API keys in the llm_cfg dictionary. The example is currently set up for a local Ollama endpoint.

Run the agent script:

# Run with direct prompt
python agentic_search.py --model "qwen3:32b" prompt "Your prompt here"

# Run with prompt from stdin
echo "Your prompt" | python agentic_search.py prompt -

# Run with custom server and API key
python agentic_search.py \
    --model "hf.co/unsloth/Qwen3-30B-A3B-GGUF:Q5_K_M" \
    --server "https://api.example.com/v1" \
    --api-key "your-key" \
    prompt "Your prompt"

Dependencies

Key Python libraries used:

qwen-agent: For creating and managing the AI agent.
mlx / mlx-lm: Likely used for efficient model inference, especially on Apple Silicon.
mcp: For integrating external tools via the Multi-Agent Collaboration Protocol.
python-dotenv: For managing environment variables (e.g., API keys).
rich: For beautiful terminal formatting and progress indicators.

See pyproject.toml for the full list of dependencies.

Contributing

This an architectural spike, I welcome your feedback through qwan.eu/contact.

License

Apache 2.0