Mano-P 1.0 - Minglue Technology's Open-Source GUI-VLA Agent Model

What Is Mano-P 1.0

Mano-P 1.0 is an open-source GUI-VLA agent model from Minglue Technology. It is purely vision-driven, requires no API, and can directly control desktop software and web interfaces. The model offers a 72B full version and a 4B quantized version, supports local deployment on Apple M4 chips, and achieves zero data upload to the cloud with physically isolated-level privacy protection. Mano-P 1.0 achieved SOTA results on 13 international benchmarks including OSWorld, is open-sourced under the Apache 2.0 license, and supports commercial applications and secondary development.

Mano-P 1.0's Main Features

• Full-Chain GUI Control: The model has complete capabilities for perception, understanding, planning, operation, and verification. It can directly control desktop software, web interfaces, and complex graphical workflows, and supports closed-loop actions such as clicking, text input, window switching, and visual verification.
• Pure Visual Understanding: It does not rely on underlying APIs, the CDP protocol, or HTML parsing. Through pixel-level visual understanding, it directly "understands" screen content, breaking the boundaries of the traditional browser ecosystem and handling non-standard applications, 3D software, and cross-system collaboration scenarios.
• On-Device Local Deployment: It supports local operation on Apple M4 chip devices (Mac mini/MacBook) and can also connect to a compute stick via USB 4.0. With zero data upload to the cloud, it achieves physically isolated-level privacy protection and can still autonomously execute long tasks in offline environments.
• Agent Ecosystem Integration:As a Skill, it seamlessly integrates into AI Agents such as OpenClaw and Claude Code, providing them with a GUI execution capability foundation and solving the bottleneck of manual intervention in complex workflows.

Mano-P 1.0's Technical Principles

• GUI-VLA Architecture: Based on a vision-language-action multimodal framework, the model directly parses screen pixel information and combines natural language instructions to output specific operation coordinates and actions, enabling cross-platform control of any graphical interface without relying on API or HTML parsing.
• Three-Stage Progressive Training: It uses supervised fine-tuning to establish foundational capabilities, optimizes the policy through offline reinforcement learning, and finally achieves real-time environmental feedback and dynamic error correction through online reinforcement learning, forming a closed-loop optimization from perception to execution.
• GSPruning Pruning Acceleration: Through proprietary visual Token pruning technology to compress redundant visual information, combined with a 4-bit quantization scheme, the 4B model achieves an inference speed of 476 tokens/s on-device on the M4 chip, with memory usage of only 4.3GB.
• Dual Device-Cloud Version Design: The 72B full model is deployed in the cloud to handle complex tasks, while the 4B quantized model focuses on local on-device operation. Combined with long-context understanding capabilities, it supports autonomous task planning and multi-step decision-making in offline environments.

How to Use Mano-P 1.0

• Get the Code: Visit the GitHub repository to clone the project source code and documentation.
• Choose a Mode: Currently, the Mano-CUA Skill can be configured for OpenClaw or Claude Code.
• Configure Integration: Connect the Skill to the target Agent so that the model gains cross-application GUI perception and automated operation capabilities.
• Local Deployment: Run the 4B quantized model on an Apple M4 chip device (32GB+ memory) to achieve offline operation with zero data upload to the cloud.
• Start Using: Drive the AI with natural language instructions to automatically parse the screen and complete complex workflows such as clicking, input, and window switching.

Mano-P 1.0's Key Information and Usage Requirements

• Product Positioning: Mano-P 1.0 is an open-source GUI-VLA agent model from MiningLamp Technology that directly controls desktop software and web interfaces through pure visual understanding, without relying on API interfaces.
• Open-source license: Fully open-source under the Apache 2.0 license, with complete auditable code, supporting commercial use and secondary development.
• Model versions: Offers a dual-version architecture with a 72B full model (high-performance cloud) and a 4B quantized model (local on-device deployment).
• Performance: Achieved SOTA results on 13 authoritative international benchmarks including OSWorld and ScreenSpot-V2, with a 58.2% task success rate on OSWorld.
• Core advantages: Pure vision-driven operation breaks the boundaries of traditional automation and supports cross-application workflows and complex graphical interface operations.
• Hardware requirements: Local deployment requires an Apple M4 chip or above with 32GB or more of memory, or connection to a Mano-P compute stick via USB 4.0.

Core advantages of Mano-P 1.0

• Pure vision-driven: No need for APIs, HTML, or underlying protocols; directly controls any desktop software and 3D applications through pixel-level understanding, breaking the boundaries of traditional automation.
• On-device local deployment: Supports local operation on devices with Apple M4 chips; the 4B quantized model requires only 4.3GB of memory, achieving physically isolated-grade privacy protection with zero data going to the cloud.
• Offline autonomous capability: Can autonomously plan and execute complex long tasks in network-free environments, with real-time decision-making and self-correction capabilities.
• Performance benchmark: The 72B version achieved SOTA on 13 international benchmarks including OSWorld, with a 58.2% task success rate on OSWorld, leading similar models by 13.2 percentage points.
• Open-source ecosystem: Fully open-source under the Apache 2.0 license, with complete auditable code, supporting commercial applications and secondary development, and seamless integration into agent ecosystems such as OpenClaw and Claude Code.

Project address for Mano-P 1.0

• GitHub repository：https://github.com/Mininglamp-AI/Mano-P

Comparison of Mano-P 1.0 with similar competitors

Application scenarios of Mano-P 1.0

• Cross-application office automation: Automatically migrate data among Excel, ERP, and email clients to complete cross-system office workflows such as report generation and email distribution.
• Complex software operation: Directly operate professional design software such as Photoshop, CAD, and 3D modeling, as well as legacy systems without API interfaces.
• End-to-end software testing: Automatically execute UI clicks, form filling, and result verification to complete full-process application testing without manual intervention.
• Privacy-sensitive business processing: Run locally to process sensitive data such as financial reports and medical records, ensuring information does not leave the local machine and meeting compliance requirements.
• Offline scenario automation: Autonomously complete long-cycle complex tasks such as data entry, document processing, and system operations and maintenance in network-free environments.