Hcompany/Holo1.5-72B
Holo1.5-72B by H Company is a 72 billion parameter Vision-Language Model (VLM) specifically designed as a foundational model for Computer Use (CU) agents, capable of interacting with real applications. Fine-tuned from Qwen/Qwen2.5-VL-72B-Instruct, it excels at UI localization and UI-based question answering across web, desktop, and mobile environments. The model achieves state-of-the-art performance on various UI localization and QA benchmarks, offering high-resolution interpretation (up to 3840 × 2160 pixels) for complex screens.
Loading preview...
Holo1.5-72B: Foundational Model for Computer Use Agents
Holo1.5-72B, developed by H Company, is a 72 billion parameter Vision-Language Model (VLM) engineered to power Computer Use (CU) agents. These agents are designed to interact with real-world applications (web, desktop, mobile) by navigating interfaces, manipulating elements, and answering content-related questions. This model is fine-tuned from Qwen/Qwen2.5-VL-72B-Instruct and is part of the Holo1.5 series, which focuses on enabling powerful automation and productivity tools.
Key Capabilities
- State-of-the-Art UI Localization: Holo1.5-72B demonstrates leading performance in accurately identifying the positions of UI elements (buttons, text boxes, etc.) on screens. It achieves an average 4.5% improvement in localization accuracy over previous models on benchmarks like Screenspot-V2, Screenspot-Pro, GroundUI-Web, Showdown, and WebClick.
- Superior UI-based Question Answering (QA): The model excels at comprehending UI structure and functionality, improving on state-of-the-art models by 3.9% in GUI QA tasks. It performs strongly on benchmarks such as ScreenQA Short, ScreenQA Complex, VisualWebBench, and WebSRC.
- High-Resolution Interpretation: Holo1.5 models are natively capable of interpreting UIs at high resolutions, up to 3840 \u00D7 2160 pixels, allowing for accurate interaction with large and complex screens.
- Advanced Training Strategy: The model is trained using a multi-stage pipeline involving high-quality proprietary data, large-scale supervised fine-tuning, and online reinforcement learning (GRPO).
Good for
- Developing next-generation Computer Use (CU) agents that can reliably interact with and automate tasks across diverse applications.
- Applications requiring precise UI element localization and robust UI content understanding.
- Research and development in AI agents for digital task delegation and productivity tools (under a research-only license).