H Company introduces Holo2, an AI model for interfaces

H Company has introduced Holo2-235B-A22B Preview, an AI model designed to locate elements within digital interfaces more accurately. The company says it reaches 78.5% on ScreenSpot-Pro and 79.0% on OSWorld G, two tests focused on whether a system can correctly identify buttons, menus, and other on-screen elements.

This is not about translating an application into another language. In this context, locating an interface means identifying exactly where a visual element is so an AI agent can interact with it: clicking a button, opening a tab, or selecting an option.

The model is already available as a research release on Hugging Face. It is the largest version of the Holo2 family so far and arrives two months after the series' first models.

How it improves screen localization

Modern interfaces can be displayed at 4K resolutions, with many small elements spread across a large screen. For an automated system, distinguishing a tiny icon from similar ones can be difficult, even when the action seems obvious to a person.

Holo2 attempts to solve this with what H Company calls agentic localization. Instead of making a single prediction, the model can review its result and refine it step by step. Each iteration helps correct the detected position and move closer to the right element.

According to the company, this process delivers relative improvements of between 10% and 20% across the different Holo2 sizes. The model does not have to get it right on the first try: it can analyze the screen again before acting.

What the tests show

On ScreenSpot-Pro, one of the more demanding benchmarks for testing whether an AI can connect instructions with graphical elements, Holo2-235B-A22B Preview achieves:

• 70.6% accuracy in a single step.
• 78.5% in up to three steps using agentic mode.

The difference shows the cost and benefit of allowing the model to review its response. It needs more time and more operations, but it locates small or difficult-to-distinguish elements more accurately.

The company also reports a result of 79.0% on OSWorld G, another test focused on using real computer environments. These figures are benchmark results, not a guarantee that the model will operate any application without errors.

For you, the potential change lies in agents that use a computer like a person: assistants capable of booking an appointment, completing a form, or working with programs that do not offer a dedicated integration. The better they can identify each button and field, the less they will depend on instructions prepared by developers.

Holo2 remains a Preview version and a research release. The important question now is whether its results hold up outside controlled tests, with changing interfaces, pop-up windows, and applications that were not designed with AI agents in mind.