Agentic AI
H Company launches Holo2-235B-A22B, breaking UI localization records
The new model achieves state-of-the-art performance on UI grounding benchmarks through iterative agentic localization.
Hugging Face Blog
H Company has released Holo2-235B-A22B Preview, its largest UI localization model to date, setting new benchmarks for screen understanding tasks.
The model achieves 78.5% accuracy on Screenspot-Pro and 79.0% on OSWorld G—state-of-the-art results for GUI grounding. Available on Hugging Face, the research release focuses specifically on UI element localization.
Agentic Localization Breakthrough#
High-resolution 4K interfaces pose a particular challenge: small UI elements can be difficult to pinpoint on large displays. H Company's approach uses iterative refinement, where the model can step through predictions to improve accuracy. This "agentic localization" unlocks 10-20% relative gains across all Holo2 model sizes.
In practice, Holo2-235B-A22B reaches 70.6% accuracy on ScreenSpot-Pro in a single step. When operating in agent mode—refining predictions over three steps—it jumps to 78.5% accuracy on the benchmark.
Infrastructure at Scale#
Training large models across multiple cloud environments requires careful orchestration. H Company uses SkyPilot to coordinate workloads, abstracting away infrastructure complexity and letting researchers focus on model development rather than managing Kubernetes manifests or deployment scripts.
