H Company has released Holo4, a family of open-weight agents that can work through a graphical interface, code, MCP tools, and APIs. The release includes a 27B dense model, a 35B-A3B mixture-of-experts model, and a smaller Holotron4 Nano. Both main models are available through H Company’s API and as downloadable weights.
The two main models have very different published desktop results. In H Company’s OSWorld 2.0 figures, Holo4 27B completes 61.7% of tasks, while Holo4 35B-A3B completes 30.9%. That is a 30.8 percentage point gap. H Company also lists Claude Opus 5.5 at 81.8% on the benchmark, although it cautions that the releases and testing setups behind comparison points differ.
One agent across several interfaces
H Company says the same Holo4 model can click and type in desktop or mobile apps, run code, and call tools when an API is available. It trained the models on supervised examples and reinforcement learning, including tasks produced by an internal system that generates interactive environments. The company says this system has made about 10,000 tasks across web apps, desktop software, and MCP servers.
The release includes public trajectories for benchmark runs, so developers can inspect the steps behind a score. H Company also shows longer examples in FreeCAD and Godot, but those demonstrations use different tasks and token counts. They are useful examples of what the agent attempts, not a controlled comparison of which model is cheaper to run.
