Holo4 series released for generalist computer-use agents with 27B and 35B-A3B variants

Hugging Face has released Holo4, a new series of agentic models designed to power generalist computer-use agents. The series includes two variants: a 27B dense model and a 35B-A3B Mixture of Experts model. Both are available via the H Models API.

Holo4 models are engineered to interact with software through multiple interfaces, including GUIs, code execution, MCP, and APIs. This design aims to address the limitations of models trained for a single interface, which may struggle in real-world workflows requiring a combination of approaches. The models are intended to run consistently across desktops, the web, Android, code sandboxes, and business APIs without needing platform-specific adaptations.

On the OSWorld 2.0 benchmark, Holo4 27B achieves a score of 61.7%, while Holo4 35B-A3B reaches 30.9%. These results place Holo4 behind only the strongest closed models, but with significantly fewer parameters and lower cost per task. The company notes that Holo4 competes on cost-performance grounds, estimating costs based on input and output tokens at H Models API rates. For context, the frontier closed models are reported to score 81.8% on the same benchmark.

Holo4 was trained using a combination of supervised and reinforcement learning on a large set of environments and tasks, including those generated by the Agentic Task Factory. This internal system produces interactive environments and verifiable tasks from documentation alone, such as screenshots of real websites or open-source software. To date, the Agentic Task Factory has generated approximately 10,000 tasks spanning web apps, MCP servers, and desktop environments.

Alongside Holo4, Hugging Face is releasing an updated version of Holotron4 Nano. The blog post provides links to the models, a full collection in multiple formats (FP16, FP8, GGUF), trajectories viewer and dataset, a quickstart guide for the H Models API, and the full blog post for further details.