Hcompany has released Holo4, two AI models designed to work across on-screen controls, code and application programming interfaces (APIs), which let software call functions directly. Both versions are available through Hcompany’s hosted API, and their weights are posted on Hugging Face.
A runtime sends screenshots and tool results to the model, then executes its requested clicks, typing, code or API calls. Hcompany lists desktop, web, Android and code-sandbox tasks among the target environments.
The two sizes have different reuse terms. Hugging Face lists Holo4-27B under CC BY-NC 4.0, which limits use to noncommercial purposes, while Hcompany says Holo4-35B-A3B is under Apache 2.0. Downloadable weights therefore do not carry the same rights across both versions.
OSWorld 2.0 awards partial credit for verified, task-specific checkpoints, so an agent can earn credit for completing parts of a workflow; full success requires satisfying every scored checkpoint. In Hcompany’s test, Holo4-27B scored 61.7% on partial credit, completed 41.5% of tasks, and cost $1.22 per task. Holo4-35B-A3B scored 30.9%, completed 12.3%, and cost $0.61 per task. The cheaper model did less of the tested work in this comparison. Hcompany says these are single runs; other systems in its table used different task subsets, test setups or effort settings, so the figures do not establish a controlled ranking.
Hcompany has also published trajectories for its benchmark runs, letting developers inspect the action sequences behind the scores. Independent repeat runs and varied workplace tasks will show whether the shared model-and-runtime approach transfers beyond the published benchmarks.
Sources
- Hcompany: Holo4 announcement, benchmark details and 35B-A3B license
- Hcompany: Holo4-27B model card and license
- OSWorld 2.0 paper: checkpoint-based partial rewards and full-completion definition
- Illustrative desktop image: SankalpSasnur, Wikimedia Commons, CC0 1.0. The photograph shows a generic computer setup, not a Holo4 device or test.
