Holo4 agent models drive screens, code and APIs from a single model
Victoria Shinder2 min readH Company has released Holo4, a new series of agentic models built to operate software through whichever interface a task needs: clicking and typing on a screen, writing and running code, or calling tools over MCP and APIs. The series comes in two sizes, a 27-billion-parameter dense model and a 35B-A3B mixture-of-experts model, both served through the company's H Models API. An updated small model, Holotron4 Nano, ships alongside.
The company's argument is that agent models are usually trained for one interface. A model built for graphical interfaces has nothing to work with when there is no screen, and a model built for tool calling is stuck in front of an application without an API. Holo4 is the same model on desktops, the web, Android, a code sandbox and business APIs, called the same way in each case.
On OSWorld 2.0, the desktop-control benchmark, Holo4 27B scores 61.7% and the mixture-of-experts model 30.9%. The strongest closed model in H Company's comparison, Opus 5.5, scores 81.8%. H Company says Holo4 trails only the strongest closed models on long workflows while costing far less per task, and it publishes every trajectory behind its public benchmark scores so each step can be replayed. The models were trained with supervised and reinforcement learning on environments generated by what the company calls its Agentic Task Factory, and both are fine-tuned from Qwen base models.
What it means
The number to read is not the gap to the leader but the cost column next to it. For workflows that must run thousands of times, a model at 61.7% that is cheap enough to retry may do more useful work than a stronger one priced per attempt. The published trajectories matter too: a benchmark score that can be replayed step by step is easier to trust than one that cannot, and the company notes that its cost figures come from its own harness and that releases and task subsets differ between the models compared.