Most agent models ship with one preferred limb: a screen, a shell, or a set of tool calls. H Company’s Holo4 arrives with all three and no favored route, keeping a single set of weights underneath.
Two sizes sit on the company’s H Models API, a 27-billion-parameter dense build and a 35B mixture-of-experts version with 3B active parameters. A smaller sibling, Holotron4 Nano, launched alongside them.
That pluralism is the pitch. A GUI-only system is helpless without a screen, and a tool-calling model has nothing to grab when an application exposes no API. Holo4 types and clicks, writes and executes its own code, and reaches for MCP or API tools, choosing whatever the task allows. The same weights answer to a desktop, a browser, an Android handset, a code sandbox and a business API, with no per-platform variant to choose.
On OSWorld 2.0, which measures desktop control, the 27B build scores 61.7 percent against 81.8 percent for Anthropic’s Opus 5.5; the mixture version reaches 30.9 percent. H Company argues those numbers undersell the model, since it runs on far fewer parameters and at much lower cost per task.
Training drew on both supervised and reinforcement learning across environments and checkable tasks produced by an in-house pipeline, the Agentic Task Factory, which assembles them from documentation alone. Every trajectory behind the published scores is open, so each run can be replayed step by step.
The demo tasks are practical and odd in equal measure: a FreeCAD model of the Eiffel Tower held to millimeter tolerances, and a self-playing Pac-Man clone in Godot left running with no keyboard input.
