Mano-P: An Open-Source On-Device AI Agent for macOS
Mano-P is an open-source, on-device GUI agent for macOS developed by Mininglamp Technology. It uses a vision-language-action architecture to operate GUI applications without system APIs or CLI access.
Why it matters
Mano-P represents the growing trend of on-device AI agents, which can operate locally without cloud dependency and ensure data privacy.
Key Points
- 1Mano-P uses a cyclical think-act-verify loop for complex workflows
- 2It is optimized for edge devices with mixed-precision quantization and GS-Pruning
- 3Mano-P has achieved competitive results on public benchmarks like OSWorld and WebRetriever
- 4Mano-P keeps all data on-device and the full client code is open-source
Details
Mano-P is designed to enable every individual and organization to create their own personalized AI agent that can run locally on their own hardware. It is purely vision-driven, understanding screen content through visual models and executing operations via native OS input. To run efficiently on consumer hardware, Mano-P uses mixed-precision quantization and the GS-Pruning algorithm for visual token compression. The model can run on an Apple M4 Pro chip with 4.3 GB of peak memory. Mano-P's training process involves three stages: SFT, offline RL, and online RL, with a bidirectional self-reinforcement framework. The project is being released in three phases, with the first phase providing Mano-CUA Skills for agent enthusiasts and the second phase introducing a local model and SDK for developers with strict data security requirements.
No comments yet
Be the first to comment