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

French AI company H has unveiled two new computer use models called Holo 4, designed to tackle graphical user interfaces (GUIs) on Windows and Linux desktops. Despite LLMs excelling at answering prompts, they struggle when it comes to navigating desktop environments, a challenge that H aims to overcome with its Holo 4 family of models.

These models can handle pointing, clicking, scrolling, and typing within graphical interfaces meant for human use. Holo 4 is built on Alibaba's Qwen 3.8 27B and Qwen 3.6 35B-A3B models and demonstrates greater versatility than pure computer use models. To showcase its capabilities, Holo 4 was used to design a 3D model of the Eiffel Tower and recreate the company's logo using extruded shapes.

H has also updated its Holotron model, based on Nvidia's Nemotron 3, with similar capabilities. While H claims superior performance compared to larger frontier models from OpenAI and Google, the models may still cost more due to increased thinking tokens required for results. However, researchers, AI enthusiasts, and enterprises can run the open weights models on relatively modest hardware, such as a 24 GB Nvidia RTX 3090 in 4-bit precision.

To facilitate usage, H has developed an open-source harness called HAI-Agents and made GGUF version of the model available on Hugging Face.