Holo4 Agentic Models Released, Featuring Generalist Computer-Use Capabilities

Introduction of Holo4 Agentic Models

H Company announced the release of Holo4, a new series of agentic models. These models are available in two primary configurations: a 27B dense model and a 35B-A3B Mixture of Experts model. Additionally, an updated version of Holotron 3, named Holotron4 Nano, was also released. All new models are accessible through the H Models API.

Generalist Interaction Capabilities

Holo4 models are designed to interact with software through various interfaces, including graphical user interfaces (GUIs), code, machine control protocols (MCP), and application programming interfaces (APIs). This multi-interface capability allows the models to adapt to different task requirements, unlike many agentic models that are specialized for a single interface type. The models can click, type, write and run code, and call tools as needed for a given task.

Training and Workflow Focus

The development of Holo4 involved supervised and reinforcement learning across a wide range of environments and tasks, including those generated by the Agentic Task Factory. While the models perform well on academic benchmarks, their primary design goal is to support real-world business workflows that often require combining different interaction methods. Holo4 operates consistently across various platforms such as desktops, web, Android, code sandboxes, and business APIs, using the same model instance for each.

Performance and Cost Considerations

Holo4 models demonstrate significant improvements over their Qwen base. On the OSWorld 2.0 benchmark, Holo4 27B achieved a score of 61.7%, while Holo4 35B-A3B reached 30.9%. These scores trail some frontier models like Opus 5.5 (81.8%) on long workflows, but Holo4 achieves this with fewer parameters and at a lower operational cost. The company also open-sources trajectories for public benchmarks, allowing for replay and download.