Mininglamp-2718/Mano-CUA-2.0-4B

VISIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Mano-CUA-2.0-4B by Mininglamp-2718 is a 4 billion parameter GUI-VLA (Visual Language Agent) model designed for autonomous desktop GUI operations. This full-precision version is part of the Mano open-source series, optimized for visual understanding and complex task automation on edge devices. It excels at cross-system data integration, long-task planning, and intelligent report generation through visual interaction.

Loading preview...

Mano-CUA-2.0-4B: Edge-Optimized GUI Automation

Mano-CUA-2.0-4B is a 4 billion parameter Visual Language Agent (VLA) model developed by Mininglamp-2718, specifically engineered for autonomous desktop GUI operations on edge devices. It is the full-precision version within the Mano open-source model series, with an 8-bit quantized MLX version also available for Apple Silicon.

Key Capabilities

  • Complex GUI Automation: Executes intricate interface operations involving hundreds of interactive elements.
  • Cross-System Data Integration: Extracts and integrates data from multiple sources purely through visual interaction, bypassing the need for APIs.
  • Long-Task Planning: Supports enterprise-level business process automation, handling workflows with dozens to hundreds of steps.
  • Intelligent Report Generation: Automatically creates structured documents like data analysis reports and work summaries.

Technical Background

The model leverages the Mano project's technical framework, incorporating a bidirectional self-reinforcement learning method, a three-stage progressive training approach (SFT → Offline RL → Online RL), and a "think-act-verify" loop reasoning mechanism. Optimizations for edge deployment include mixed-precision quantization, visual token pruning, and edge inference adaptation, enabling efficient execution on devices like Mac mini/MacBook.

Action Space

Mano-CUA-2.0-4B provides a comprehensive action space for GUI interaction, including open_app, open_url, various click types (single, double, triple, right), hover, type, hotkey, scroll, drag, wait, finish, stop, and call_user for human assistance. The model outputs structured XML actions with normalized coordinates.

Finetunes
2 models