H Company ships Holo2-235B A22B Preview for higher-accuracy UI element localization

UI automation lives or dies on whether a model can reliably point to the right pixel. H Company is betting that better grounding—accurately locating buttons, fields, and menus on real screens—will unlock more dependable AI agents for testing, customer support, and back-office ops.

New model targets UI grounding benchmarks

H Company has released Holo2-235B-A22B Preview, a research model focused on UI element localization and published it on Hugging Face. In results shared by the company, the model sets a new top score on the ScreenSpot-Pro GUI grounding benchmark, reaching 78.5 percent accuracy, and records 79.0 percent on OSWorld G.

For marketers and e-commerce operators, this category of model matters because it underpins “do this in the app” workflows: updating product listings, configuring ad dashboards, pulling reports, or validating that a landing page renders correctly. When localization fails, automations misclick and break. When it works, teams can push more processes into scripts and agents.

Agentic localization improves accuracy on high-resolution UIs

A key claim in the release is that the model benefits from “agentic localization,” meaning it can take multiple passes to refine where an on-screen target is located. H Company says this is especially useful on 4K interfaces where small UI controls are easy to miss.

On ScreenSpot-Pro, Holo2-235B-A22B Preview reportedly scores 70.6 percent in a single step, then climbs to 78.5 percent within three steps in agent mode—representing a 10 to 20 percent relative gain across Holo2 sizes, according to the company. In practical terms, iterative refinement is similar to an agent “zooming in” conceptually: it proposes a location, checks, and adjusts, rather than committing to one guess.

Better localization doesn’t automatically mean full end-to-end task success, but it reduces one of the biggest failure points for UI-driven automation: reliably selecting the right element under real-world layouts.