Hcompany releases Holo1.5 with new 72B model for UI tasks

Hcompany has announced the release of Holo1.5, an upgraded series of open-source models designed for Computer Use agents, now available in 3B, 7B, and a new 72B parameter size. The release targets developers, researchers, and companies building digital agents that need precise UI element localization and robust UI understanding across various platforms, including macOS, Ubuntu, Windows, web, and mobile environments. The models are publicly accessible via Hugging Face, supporting immediate integration and experimentation.

Holo1.5 models feature a more than 10% accuracy increase over the previous Holo1 series, setting new benchmarks for UI localization and visual question answering (VQA) tasks. These models output coordinates for UI navigation and answer questions about visual states with higher accuracy than competitors like Qwen-2.5 VL, Sonnet 4, UI-TARS 1.5, and UI-Venus. Technical evaluations show strong performance in high-resolution and complex GUI environments such as Photoshop and AutoCAD, emphasizing their suitability for professional use cases.

Early industry feedback underscores the model’s improved reliability and efficiency in real-world applications. The open-weight release signals Hcompany’s intention to foster wider adoption and encourage the development of more capable cross-platform digital agents. The architecture’s scalability, with the introduction of the 72B model, allows users to select models that fit their computational resources and accuracy needs. Hcompany’s ongoing focus is to advance foundational AI for computer use, aiming to set new standards for trust and utility in digital agent technology.

Source