Hcompany released two new agentic models named Holo4

Hcompany introduced Holo4, a series of models designed to act as generalist agents that operate software just like a human would. The release includes a 27B dense model and a 35B-A3B mixture of experts model, both accessible through the H Models API.

These agents do not rely on a single input method. They can interact with graphical user interfaces, write and execute code, and use application programming interfaces or Model Context Protocol tools. This allows them to handle business workflows that span across different digital environments.

Performance benchmarks show that these models compete with frontier systems on desktop control tasks like OSWorld 2.0. While larger models maintain a lead in overall scoring, Holo4 achieves results at a significantly lower operational cost and with fewer parameters.

To improve reliability, the development team updated their internal harness to help agents maintain memory over long task sequences. They also integrated a desktop shell that allows the models to track their progress across hundreds of steps.

Alongside the main models, Hcompany also released Holotron4 Nano. This smaller version is designed to perform well on GUI workflows and environments using APIs or coding sandboxes.

Despite these advances, the current iteration of Holo4 has limitations. The team has not yet reported performance data for the private set of the AutomationBench benchmark, and the models are still being evaluated against the highest-performing closed systems in specific complex scenarios. Would you trust an AI to navigate your desktop and run code for you?