H company released Holo4, a new series of agentic models in 27B dense and 35B-A3B sizes. Both models work across GUIs, code, MCP, and APIs without switching. They are available on the H Models API and as open weights. Holo4 also includes Holotron4 Nano, an updated smaller model. It scores 61.7% and 30.9% on OSWorld 2.0 for the 27B and 35B-A3B versions. Cost per task is far lower than frontier models with similar performance. All benchmark trajectories are open-sourced for replay. Models were trained on tasks from an Agentic Task Factory. They excel on professional software like FreeCAD and Godot. Holo4 runs on desktops, web, Android, code sandboxes, and business APIs.
What changed
Previous H company agentic models were limited to one interface type per model.
What it unlocks
Using one model across GUIs, code, MCP, and APIs for the same task.
- 27B dense model
- 35B-A3B Mixture of Experts model
- 61.7% OSWorld 2.0 score for 27B
- 30.9% OSWorld 2.0 score for 35B-A3B
What you need to act on it
- access to H Models API
- self-hosting capability for local runs
Sources