QwenPaw: building AI that you can trust
QwenPaw: building AI that you can trust
- Day
- Day 2 — Session Day 1
- Time
- 2:50pm-3:10pm
- Room
- Track 1
- Track
- Claws & Personal Agents
Accessible with the Engineering pass and above.
About this session
QwenPaw (previously CoPaw) is a local-first personal assistant from Alibaba that installs with choices of 2b, 4b and 8b models, all customized to work well with common tasks from users. It is open source, available on GitHub (github.com/agentscope-ai/QwenPaw). Our goal is to build an AI assistant that you can trust your personal data with. In this talk we cover the key problems facing personal AI assistant: - How do you finetune a small model to be genuinely useful in a personal assistant setting? - How do you leverage cloud models for difficult tasks without leaking private context? We'll share the routing architecture that keeps sensitive context on-device while still leveraging cloud models for complex reasoning. Finally, we'll share the story of QwenPaw's open-source journey in China — including the "lobster frenzy" that unexpectedly ignited community contributions — and what we've learned about building an open developer community around local-first AI.
