Mano-P is an open-source GUI agent from Mininglamp Technology that operates computers through pure visual understanding and runs entirely on-device. Its 72B model tops the OSWorld benchmark for dedicated GUI agents at 58.2%, leading the runner-up by 13.2 percentage points, while a quantized 4B version (w4a16) runs locally on an M4 Mac mini with 32GB RAM, peaking at 4.3GB memory with 476 tokens/s prefill and 76 tokens/s decode. Unlike cloud-based solutions such as Claude Computer Use, Mano-P never uploads screenshots, addressing privacy, latency, and API lock-in. Key techniques include bidirectional Mano-Action self-reinforcing learning with a Think-Act-Verify loop, GSPruning visual token pruning (12.57% token retention, 2-3x throughput gain), and mixed-precision quantization. It ships in three forms: a CLI (mano-cua), a Python SDK (mano-client, upcoming), and an OpenClaw/Claude Code skill. Released under Apache 2.0 with a three-stage open-source roadmap, Mano-P is fully offline-capable, UI-agnostic, and open to commercial use.
Mano-P Deep Dive: A GUI Agent That Leaves the Cloud and Runs on Your Mac
Summary
Mano-P is an open-source GUI agent from Mininglamp Technology that operates computers through pure visual understanding and runs entirely on-device. Its 72B model tops the OSWorld benchmark for dedicated GUI agents at 58.2%, leading the runner-up by 13.2 percentage points, while a quantized 4B version (w4a16) runs locally on an M4 Mac mini with 32GB RAM, peaking at 4.3GB memory with 476 tokens/s prefill and 76 tokens/s decode. Unlike cloud-based solutions such as Claude Computer Use, Mano-P never uploads screenshots, addressing privacy, latency, and API lock-in. Key techniques include bidirectional Mano-Action self-reinforcing learning with a Think-Act-Verify loop, GSPruning visual token pruning (12.57% token retention, 2-3x throughput gain), and mixed-precision quantization. It ships in three forms: a CLI (mano-cua), a Python SDK (mano-client, upcoming), and an OpenClaw/Claude Code skill. Released under Apache 2.0 with a three-stage open-source roadmap, Mano-P is fully offline-capable, UI-agnostic, and open to commercial use.
This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618890