Hcompany has launched Holo4, a series of agentic AI models designed to operate across diverse software interfaces including GUIs, code sandboxes, and APIs. The release includes a 27B dense model and a 35B Mixture of Experts version, alongside a smaller Holotron4 Nano model for lightweight tasks.
Cross-platform task automation
The Holo4 series utilizes a single model to navigate multiple software environments, such as desktops, web browsers, and Android devices. Unlike specialized agents restricted to either visual interfaces or API tool-calling, these models are trained to combine methods, allowing them to type on screens and run code within the same workflow.
Hcompany employed a combination of supervised and reinforcement learning to develop the series. Training data was sourced from various environments and a proprietary task generation system. The models are available via the H Models API, with the developer providing open-source access to the interaction trajectories used for public benchmarking.
Performance and interface limitations
While Holo4 models are smaller than frontier counterparts, they show competitive results on benchmarks like OSWorld 2.0. The 27B dense version scored 61.7% on long workflows, trailing behind larger closed models like Opus 5.5 which reached 81.8%. The 35B Mixture of Experts version achieved a 30.9% score on the same test.
The models rely on the Model Context Protocol and standard APIs to supplement their graphical navigation capabilities. Although designed for business workflows, the reported effectiveness is based on academic benchmarks and internal task environments. Independent verification of performance in live, unmanaged enterprise settings is not included in the release.
Original source
This report summarises the source below. Analysis is labelled separately; product and research claims remain attributed to their source.
Read the original at Hugging Face