---
language:
- en
license: other
library_name: transformers
tags:
- tool-calling
- mcp
- browser-automation
- lora
- ccmcp
- mlx
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
datasets:
- custom
pipeline_tag: text-generation
---

# qwen-3b-ccmcp-v1

Qwen 3B trained for Claude Chrome MCP tool calling

> **Attribution:** Built with Qwen


## Model Description

Fine-tuned for MCP (Model Context Protocol) tool calling with the Claude Chrome extension. The model generates tool calls for browser automation tasks.

## Training Details

- **Base Model:** Qwen/Qwen2.5-Coder-3B-Instruct
- **Method:** LoRA fine-tuning on Apple Silicon (MLX)
- **Dataset:** 1,782 MCP browser automation examples
- **Validation Loss:** 0.077
- **Iterations:** 1000
- **Naming Convention:** `{base}-{size}-ccmcp-{version}`
  - `ccmcp` = **C**laude **C**hrome **MCP**

## Files

- `adapters.safetensors` - LoRA adapter weights
- `adapter_config.json` - LoRA configuration
- `qwen-3b-ccmcp-v1-f16.gguf` - GGUF F16 format for llama.cpp/Ollama
- `checkpoints/` - Training checkpoints

## Usage

### With MLX (Apple Silicon)

```python
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load(
    "mlx-community/Qwen2.5-Coder-3B-Instruct-4bit",
    adapter_path="pierretokns/qwen-3b-ccmcp-v1"
)

messages = [
    {"role": "system", "content": "You are a browser automation assistant with MCP tools."},
    {"role": "user", "content": "Go to google.com"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
sampler = make_sampler(temp=0.1)

response = generate(model, tokenizer, prompt=prompt, max_tokens=150, sampler=sampler)
print(response)
```

### With Ollama

```bash
# Download GGUF from this repo
# Create Modelfile:
cat > Modelfile << 'EOF'
FROM ./qwen-3b-ccmcp-v1-f16.gguf
PARAMETER num_ctx 8192
PARAMETER temperature 0.1
SYSTEM "You are a browser automation assistant with MCP tools."
EOF

# Create and run
ollama create qwen-3b-ccmcp-v1 -f Modelfile
ollama run qwen-3b-ccmcp-v1 "Go to google.com"
```

### With Claude Code + Ollama

```bash
ANTHROPIC_BASE_URL=http://localhost:11434 \
ANTHROPIC_AUTH_TOKEN=ollama \
ANTHROPIC_API_KEY=ollama \
claude --model qwen-3b-ccmcp-v1
```

## MCP Tools

The model was trained on 16 MCP browser automation tools:
navigate, read_page, find, computer, form_input, get_page_text, screenshot,
javascript_tool, tabs_context_mcp, tabs_create_mcp, gif_creator, upload_image,
read_console_messages, read_network_requests, shortcuts_list, shortcuts_execute

## License

This model inherits the license from the base model. See [Qwen/Qwen2.5-Coder-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct) for license details.

Full license: [https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/blob/main/LICENSE](https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/blob/main/LICENSE)