A key point of Holo3 is that it aims not only to recognize screens but also to carry out tasks across multiple apps. [Photo: H Company]

French artificial intelligence startup H Company has unveiled Holo3, an AI model that can read screens and carry out clicks, inputs and tasks across apps.

Online media outlet Gigazine reported on April 9 local time that the open-source version, Holo3-35B-A3B, is available for free on Hugging Face.

Holo3 is a large vision-language model that runs in web, desktop and mobile environments. It is designed to read screen information and perform context-appropriate actions such as pressing buttons or filling out forms.

The model goes beyond simple click automation to handle tasks that move across multiple apps. For example, it can extract equipment price information from a PDF file, compare it with remaining budgets by employee, and send an approval or rejection email. It can move between PDFs, spreadsheets and email to read documents, do calculations and relay information, then keep the task state and continue to the next step.

The open-source Holo3-35B-A3B is fine-tuned from Qwen3.5-35B-A3B. Using a mixture-of-experts structure, it has 35 billion total parameters and 3 billion active parameters used in operation. It is built as a multimodal AI that takes images and text as inputs and generates text.

Training used open-source datasets, large-scale interaction data made for AI, and human-reviewed annotated data. It was trained to respond more easily to situations not used in training, and combined with selected reinforcement learning. The company also set up a Synthetic Environment Factory that automatically builds a UI and interaction environment close to enterprise systems with a code-generation agent, and used it to train operations similar to work tasks.

It also released performance results. Holo3-35B-A3B scored 77.8 percent on the international standard benchmark OSWorld-Verified. The higher-end Holo3-122B-A10B scored 78.85 percent on the same benchmark. It has 122 billion total parameters and 10 billion active parameters.

H Company also presented its in-house benchmark, the H Corporate Benchmark, made up of 486 tasks across 4 areas: e-commerce, business software, collaboration and multi-app integration. It includes everything from short tasks that finish within a single app to longer workflows that span multiple apps.

In the free tier, users can try Holo3-35B-A3B via an API with a request limit of 10 per minute. The higher-end Holo3-122B-A10B is available only in the paid tier.

Keyword

#H Company #Holo3 #Hugging Face #OSWorld-Verified #Qwen3.5-35B-A3B
Copyright © DigitalToday. All rights reserved. Unauthorized reproduction and redistribution are prohibited.