H Company has released Holo4, a pair of computer-use models designed to move between graphical interfaces, code, Model Context Protocol tools, and APIs within the same task. The models come as a 27-billion-parameter dense version and a 35B-A3B mixture-of-experts version, and both are available through the H Models API.
The company says Holo4 was trained with supervised and reinforcement learning across desktop, web, Android, sandboxed code, and business API environments. This matters because many agents are optimized either for clicking through a screen or for structured tool calls; real workflows often require both. H Company also published the trajectories behind its public benchmark scores so researchers can replay the steps.
On OSWorld 2.0, a difficult desktop-control benchmark, the company reports a 61.7% score for Holo4 27B and 30.9% for the 35B-A3B model, compared with 81.8% for Opus 5.5. Those are vendor-reported benchmark results, not a guarantee of reliability in unsupervised business use. An updated Holotron4 Nano model is also being released, while access to the main Holo4 models currently centers on the company’s hosted API.