Mano-P 1.0 - An open-source GUI-VLA intelligent agent model from Minglue Technology.

What is Mano-P 1.0?

Mano-P 1.0 is an open-source GUI-VLA intelligent agent model from Minglue Technology. It is purely vision-driven and can directly control desktop software and web interfaces without requiring an API. The model provides a 72-byte full version and a 4-byte quantized version, supports local deployment on Apple M4 chips, and achieves zero-cloud data transfer and physical isolation-level privacy protection. Mano-P 1.0 has achieved state-of-the-art (SOTA) results in 13 international benchmark tests, including OSWorld, and is open-source under the Apache 2.0 license, supporting commercial applications and secondary development.

Main features of Mano-P 1.0

• GUI end-to-end controlThe model possesses 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 understandingIt does not rely on underlying APIs, CDP protocols, or HTML parsing. It can directly "understand" screen content through pixel-level visual understanding, breaking the boundaries of the traditional browser ecosystem and handling non-standard applications, 3D software, and cross-system collaboration scenarios.
• Local deployment on the device sideSupports local operation on Apple M4 chip devices (Mac mini/MacBook), and can also be connected to a computing stick via USB 4.0. Zero data upload to the cloud, achieving physical isolation-level privacy protection, and can still autonomously execute long-duration tasks in offline environments.
• Agent ecosystem integration:Skill seamlessly integrates with AI agents such as OpenClaw and Claude Code, providing them with a GUI execution capability foundation to solve the bottleneck of manual intervention in complex workflows.

Technical principles of Mano-P 1.0

• GUI-VLA ArchitectureBased on a vision-language-action multimodal framework, the model directly parses screen pixel information and outputs specific operation coordinates and actions by combining natural language commands. It can control any graphical interface across platforms without relying on API or HTML parsing.
• Three-stage progressive trainingThe system employs supervised fine-tuning to lay the foundation for capabilities, optimizes strategies 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 AccelerationBy compressing redundant visual information through proprietary visual token pruning technology and combining it with a 4-bit quantization scheme, the 4B model achieves an inference speed of 476 tokens/s on the M4 chip side with a memory footprint of only 4.3GB.
• Dual-version design for endpoints and cloudThe 72B full model is deployed in the cloud to handle complex tasks, while the 4B quantized model focuses on local operation on the device side. 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 codeAccess the GitHub repository to clone the project's source code and documentation.
• Select modeCurrently, Mano-CUA Skill can be configured to OpenClaw or Claude Code.
• Configuration integrationConnect Skills to the target Agent to enable the model to gain cross-application GUI awareness and automated operation capabilities.
• Local deploymentRun the 4B quantization model on an Apple M4 chip device (32GB+ memory) to achieve offline operations with zero data upload to the cloud.
• Start usingIt uses natural language commands to drive AI to automatically analyze the screen and complete complex workflows such as clicking, inputting, and switching windows.

Key information and usage requirements of Mano-P 1.0

• Product PositioningMano-P 1.0 is an open-source GUI-VLA intelligent agent model from Minglue Technology. It can directly control desktop software and web interfaces through pure visual understanding without relying on API interfaces.
• Open source licenseIt is fully open source under the Apache 2.0 license, with complete auditable code, and supports commercial use and secondary development.
• Model versionIt offers a dual-version architecture: a 72B full model (high performance in the cloud) and a 4B quantized model (local deployment on the device).
• PerformanceIt achieved state-of-the-art (SOTA) results in 13 international authoritative benchmark tests, including OSWorld and ScreenSpot-V2, with a success rate of 58.2% in OSWorld tasks.
• Core advantagesPure vision-driven operation breaks the boundaries of traditional automation and supports cross-application workflows and complex graphical interface operations.
• Hardware configurationLocal deployment requires an Apple M4 chip or higher device with at least 32GB of memory, or a Mano-P computing stick connected via USB 4.0.

The core advantages of Mano-P 1.0

• pure vision-drivenIt requires no API, HTML, or underlying protocols to directly control any desktop software and 3D application through pixel-level understanding, breaking the boundaries of traditional automation.
• Local deployment on the device sideSupports local operation on Apple M4 chip devices; the 4B quantization model requires only 4.3GB of memory, achieving physical isolation-level privacy protection with zero data upload to the cloud.
• Offline autonomyIt can autonomously plan and execute complex and long tasks in environments without network access, and has the ability to make real-time decisions and self-correct.
• Performance benchmarkVersion 72B achieved state-of-the-art (SOTA) results in 13 international benchmark tests, including OSWorld, with a success rate of 58.2% on OSWorld tasks, leading similar models by 13.2 percentage points.
• Open source ecosystemThe Apache 2.0 license is fully open source, the complete code is auditable, it supports commercial applications and secondary development, and seamlessly integrates with the OpenClaw, Claude Code and other agent ecosystems.

The project address for Mano-P 1.0

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

Comparison of Mano-P 1.0 with similar competing products

Application scenarios of Mano-P 1.0

• Cross-application office automationAutomatically migrates data between Excel, ERP, and email clients, completing cross-system workflows such as report generation and email distribution.
• Complex software controlDirectly control professional design software such as Photoshop, CAD, and 3D modeling, as well as legacy systems without API interfaces.
• End-to-end software testingIt automatically executes UI interface clicks, form filling, and result verification, completing the entire application testing process without human intervention.
• Privacy-sensitive business processingIt processes sensitive data such as financial statements and medical records locally, ensuring that the information does not leave the local machine and meeting compliance requirements.
• Offline scene automationIt can autonomously complete long-cycle and complex tasks such as data entry, document processing, and system maintenance in environments without network access.