H Company Holo2-235B-A22B Preview Release

H Company has released the Holo2-235B-A22B Preview, a large-scale research model specifically designed for UI element localization. The model establishes new state-of-the-art (SOTA) records on the Screenspot-Pro benchmark (78.5%) and the OSWorld G benchmark (79.0%).

UI Localization Performance and Benchmarks

Holo2-235B-A22B Preview delivers industry-leading accuracy in pinpointing UI elements. In a single-step prediction, the model reaches 70.6% accuracy on ScreenSpot-Pro. When operating in agent mode, it achieves 78.5% accuracy within three steps, setting a new SOTA record for the GUI grounding benchmark.

Agentic Localization for High-Resolution Interfaces

Agentic localization allows Holo2 to iteratively refine its predictions to overcome the challenges of high-resolution 4K interfaces, where small UI elements are often difficult to locate. This iterative approach provides 10-20% relative gains in accuracy across all Holo2 model sizes.

Infrastructure and Training Scale

To manage the scale of training Holo2, H Company utilizes SkyPilot. This tool serves as a unified interface for launching training jobs across multiple cloud providers using Kubernetes (k8s). By using SkyPilot to abstract infrastructure complexity, researchers can focus on model development rather than managing k8s manifests or deployment scripts.