French dev aims to solve bots' blindness so they can understand GUIs

French AI model developer H has unveiled a pair of computer use models named Holo 4 and Holotron, designed to help machines understand graphical user interfaces (GUIs) in Windows and Linux operating systems. The Holo 4 family of models aims to bridge the gap between command line interfaces (CLIs), application programming interfaces (APIs), and GUIs, which have traditionally been the three main categories of computer interaction.

Holo 4 enables relatively small but powerful models to perform tasks such as pointing, clicking, scrolling, and typing within graphical interfaces meant for human users.

Developed on top of Alibaba's Qwen 3.8 27B and Qwen 3.6 35B-A3B models, Holo 4 has been fine-tuned using supervised training and reinforcement learning methods. This combination results in a versatile model with greater capability compared to pure computer use models. To improve performance, H has also updated its Holotron model, which is based on Nvidia's Nemotron 3.

In demonstration, Holo 4 27B utilized FreeCAD's macro function to programmatically design a 3D model of the Eiffel Tower, showcasing its ability to work with graphical interfaces. Additionally, the model was able to recreate H's own logo using extruded shapes, illustrating its flexibility and capabilities.

Although Holo 4 outperforms larger frontier models from companies like OpenAI while using fewer parameters, it may cost more per task. This is likely due to the model using more "thinking" tokens during execution. Despite this, researchers, AI enthusiasts, and enterprises can still run the models on relatively modest hardware, such as a 24 GB Nvidia RTX 3090 at 4-bit precision.

H has made the model available in various weight formats, including Llama.cpp (via LM Studio and Ollama), and plans to release additional DSpark draft weights for enhanced inference using speculative decoding techniques.