Holo4: powering generalist computer-use agents
Hugging Face introduces Holo4, a new series of agentic AI models designed for general computer-use tasks across multiple interfaces, available via H Models API.
Holo4 is a new series of agentic AI models from Hugging Face, available in two sizes: a 27B dense model and a 35B-A3B Mixture of Experts model. Both are accessible through the H Models API, alongside an updated version of Holotron 3 called Holotron4 Nano. The models are designed to interact with software through various interfaces, including GUIs, code, MCP, and APIs, enabling them to perform tasks such as clicking, typing, writing code, and calling tools.
Holo4 was trained using supervised and reinforcement learning on a large set of environments and tasks, including those generated by Hugging Face's Agentic Task Factory. Unlike many agentic models that specialize in a single interface, Holo4 can adapt to different platforms and workflows, running on desktops, the web, Android, in code sandboxes, and against business APIs without requiring model selection per platform.
The models demonstrate strong performance on academic benchmarks like OSWorld 2.0 and AutomationBench, competing with frontier closed models while maintaining significantly lower costs and fewer parameters. For example, Holo4 27B scores 61.7% on OSWorld 2.0 compared to 81.8% for Opus 5.5, and Holo4 35B-A3B reaches 30.9%, with costs estimated per task based on input and output tokens.
Holo4 excels in professional software tasks, as demonstrated by examples such as building a 3D model of the Eiffel Tower in FreeCAD, creating a logo in FreeCAD, and developing a Pac-Man-style game in Godot. The models were trained on approximately 10,000 tasks across web apps, MCP servers, and desktop environments, with an improved harness that provides reliable memory and desktop shell integration.