Holo2-235B-A22B Preview advances UI localization benchmarks (Screenspot-Pro 78.5%, OSWorld G 79.0%)

AI Impact Summary

H Company's Holo2-235B-A22B Preview is a 235B UI localization model released as a research artifact on Hugging Face. In agent mode it reaches 78.5% on Screenspot-Pro within 3 steps and 70.6% in a single step, with OSWorld G at 79.0%, establishing a new SOTA for GUI grounding. The emphasis on agentic localization yields 10-20% relative gains across Holo2 sizes, signaling meaningful uplift for large-scale 4K UI localization pipelines; production teams should plan evaluation against internal assets, assess latency and compute needs for multi-step inference, and prepare scalable training/inference workflows via SkyPilot and Kubernetes.