H Company’s new Holo2 model takes the lead in UI Localization

Source: HuggingFace Date: 2026-02-03 URL: https://huggingface.co/blog/Hcompany/introducing-holo2-235b-a22b

Summary

Model release: H Company’s Holo2-235B-A22B, a 235B parameter UI localization model achieving 78.5% on ScreenSpot-Pro and 79.0% on OSWorld G (new SOTA). Single-step accuracy is 70.6%; agentic localization with iterative refinement over 3 steps reaches 78.5% — a 10–20% relative gain. Trained with SkyPilot for multi-cloud Kubernetes coordination. Research preview.

Implications

Thread: open-weights ecosystem health / agentic patterns. Holo2’s agentic localization result is notable: iterative self-refinement on 4K UI screenshots yields 10–20% gains over single-step inference. This validates the pattern that GUI agents benefit from multi-step refinement loops rather than single-shot prediction — the accuracy gap narrows with compute budget. At 235B parameters this is not a local-deployment model, but the technique is transferable to smaller models. H Company’s trajectory (Holo1 → Holo2) suggests GUI agent capabilities are compounding faster than general reasoning benchmarks.