qwen3-0.6b-notifai

Qwen3-0.6B-NotifAI

Fine-tuned Qwen3-0.6B model for on-device notification classification. Part of the NotifAI project.

Model Description

This model classifies mobile notifications into folders and priority levels, running entirely on-device for privacy.

Intended Use

• On-device notification classification for Android
• Privacy-preserving notification management
• Real-time classification (1.3s latency on mobile)

Classification Output

{"folder": "Work", "priority": "high"}

Folders

• Work - Professional notifications (Slack, Teams, Jira)
• Personal - Friends & family (WeChat, WhatsApp)
• Promotions - Marketing & deals
• Alerts - System & transactional

Priority

• high - Immediate attention needed
• medium - Important but not urgent
• low - Can wait

Usage with llama.cpp

./llama-cli -m qwen3-0.6b-notifai-Q8_0.gguf -p "<prompt>"

Training

• Method: LoRA fine-tuning with Unsloth
• Hardware: RTX 4060 Ti 16GB
• Duration: 81 minutes
• Epochs: 3
• Learning Rate: 2e-4

Limitations

• Optimized for common notification types
• Best performance with English and Chinese text
• Requires ~650MB storage on device

Citation

@misc{notifai2026,
  title={NotifAI: On-Device Notification Classification},
  author={封一},
  year={2026},
  url={https://github.com/charfeng1/notifai}
}

License

Apache 2.0