Alibaba Tongyi Lab Releases CoPaw 1.0: Supports Local Deployment and Efficient Multi-Agent Collaboration

Alibaba Tongyi Lab recently released CoPaw 1.0, focusing on upgrading four core capabilities: custom small models, security mechanisms, multi-agent collaboration, and memory management. Built on the AgentScope ecosystem, it uses a three-layer architecture: framework, memory, and model layers.

The new version supports local deployment on Mac, Windows, and Linux systems, allowing users to install and run with one click. The team also launched the CoPaw-Flash series of custom small models, offering various sizes and quantized versions that auto-recommend based on device performance—reducing cloud costs while ensuring data security.

For security, CoPaw 1.0 uses a layered defense architecture, including tool guards for runtime tool call detection, file protection for sensitive path access control, and skill scanners for pre-installation risk checks. Multi-agent collaboration is enhanced: users can create isolated workspaces for different tasks, with agents running in parallel, independently configured, and supporting asynchronous communication to handle complex workflows. Memory management is driven by the ReMe mechanism, dynamically organizing dialogue history and balancing coherence with efficiency.