H Company is shaking up interface localization again: their new model Holo2-235B-A22B Preview sets a record on GUI grounding benchmarks, designed specifically to identify and locate UI elements on high-resolution screens.
What is Holo2-235B-A22B and why it matters
Holo2-235B-A22B Preview is a 235-billion-parameter model released as a research release on Hugging Face, focused on UI element localization. In public tests it reaches 78.5% on ScreenSpot-Pro and 79.0% on OSWorld G, two relevant benchmarks for grounding evaluation in interfaces.
Why does this matter to you? Locating tiny buttons, icons, and text on 4K screens is hard: there are few pixels and a lot of context. A jump in accuracy here directly affects accessibility tools, automated testing, workflow automation, and visual assistants.
Agentic localization: iterate to improve
The big technical novelty is the agentic localization mode. Instead of giving a single prediction, the model can iterate: it refines its output step by step and corrects mistakes. On ScreenSpot-Pro Holo2-235B-A22B gets to 70.6% in a single step, but in agent mode it reaches 78.5% in 3 steps. In other words, the ability to deliberate and adjust delivers substantial improvements.
Why is that useful in practice? Think of it like proofreading a screenshot: a first pass finds most things, a couple of focused passes catch the small, easy-to-miss elements.
H Company reports this strategy unlocks relative gains of 10–20% across all sizes of the Holo2 family. It’s a neat demonstration that smarter inference flows—not just blindly bigger models—can outperform parameter increases without strategy.
Technical implications and recommendations
-
Latency vs accuracy: agent improvements mean more inference steps. Is it worth it for your use case? For automated tests or assisted accessibility, likely yes; for real-time UX, measure latency.
-
Pipeline design: combine Holo2 outputs with OCR and spatial heuristics. Often a spatial detector + clustering heuristic reduces false positives.
-
Evaluation: test on ScreenSpot-Pro and OSWorld G to compare with the SOTA. Reproduce the agent steps when you evaluate, because the single-step version doesn’t show the full potential.
-
Data and fine-tuning: if you work with industrial UIs or specific apps, consider fine-tuning with synthetic examples that reflect your element density and 4K resolutions.
-
Computational costs: 235B has high requirements. For production, explore smaller models in the Holo2 family or distillation strategies if latency is critical.
Practical use cases
-
Accessibility: assistants that point out elements on screen for users with visual impairments.
-
Visual QA: automate interface tests and detect regressions in complex layouts.
-
RPA and task automation: identify and click dynamic elements in high-resolution desktop and web apps.
-
Internationalization and localization: verify layout consistency when changing language or format.
And if you’re just curious? This shows a clear trend: it’s not only about bigger models, but about inference modes that think in steps.
H Company published this model as a research release on Hugging Face so the community can evaluate and explore these ideas in real scenarios.
Original source
https://huggingface.co/blog/Hcompany/introducing-holo2-235b-a22b