Holo1: Open-Source Vision Models for GUI Automation Achieve 76.2% Accuracy

H Company has introduced Holo1, the first family of open-source Action Vision Language Models (VLMs) built specifically for GUI automation and web UI understanding. The release includes Holo1-3B and Holo1-7B models, with the larger variant achieving 76.2% average accuracy on common UI localization benchmarks—the highest performance among small-size models in its category.

Key Features#

The Holo1 models are built on the Qwen2.5-VL architecture and integrate seamlessly with the Hugging Face Transformers library. They're accompanied by the WebClick benchmark, containing 1,639 human-like UI interaction tasks for evaluation and development.

The models excel at precise element localization—identifying exactly where to click on a webpage based on natural language instructions. A practical example: given a calendar interface and the instruction "Select July 14th as the check-out date," Holo1 outputs specific pixel coordinates for the interaction.

Powering Surfer-H#

These models serve as the backbone for Surfer-H, a web-native automation agent designed to interact with browsers like a human would. Surfer-H achieves 92.2% accuracy on real-world web tasks at just $0.13 per task, offering a cost-efficient alternative to traditional automation solutions.

The agent uses a modular three-component architecture:

• Policy model: Plans and drives the agent's behavior
• Localizer model: Understands visual UIs for precise interactions
• Validator model: Confirms task completion

Unlike systems relying on custom APIs or fragile wrappers, Surfer-H operates purely through standard browser interactions, mimicking human behavior.

Availability#

Both Holo1 models and Surfer-H are available as open-source releases on Hugging Face, enabling developers to build cost-efficient web automation solutions.