AI model 'Holo3' capable of performing click operations and task execution on a PC has been released; the open-source version has a free tier.
H Company, a French AI startup, released ' Holo3 ,' an AI model capable of performing click operations and task execution on a PC, on March 31, 2026. Holo3 is a large-scale vision language model optimized for GUI agents, and the open-source version, 'Holo3-35B-A3B,' is available for free from Hugging Face, an AI model publishing and distribution service.
Holo3 - H company
Hcompany/Holo3-35B-A3B - Hugging Face
https://huggingface.co/Hcompany/Holo3-35B-A3B
Holo3 is a large-scale vision language model that operates in digital environments such as web, desktop, and mobile. It is designed to read information on the screen and perform contextual actions such as pressing buttons and filling out forms.
For example, the system is designed to handle processes involving multiple applications, such as extracting equipment pricing information from a PDF file, comparing it against each employee's remaining budget, and then sending approval or rejection emails to each individual. This goes beyond simple clicks; it includes reading and calculating documents, transferring information between multiple applications, and continuing the work while maintaining the current state of the process, all while navigating across PDF files, spreadsheets, and emails.
The open-source version of 'Holo3-35B-A3B' released this time is a finely tuned model based on ' Qwen/Qwen3.5-35B-A3B '. It employs a Mixture of Experts configuration, which runs using only a portion of multiple processing parts, with a total of 35 billion parameters, of which 3 billion are active parameters used during operation. It is a multimodal AI that takes images and text as input and generates text.
The training process utilizes open-source datasets, as well as a large amount of artificially created operation data for AI, and training data that has been reviewed and annotated by humans. By combining this data, the AI's ability to distinguish screen content and decide on the next action is enhanced. Furthermore, in addition to using artificially created operation data, the AI is trained to handle situations not used in training, and carefully selected reinforcement learning techniques are also incorporated.
A system called 'Synthetic Environment Factory' is also available, which uses a code generation agent to automatically create a UI and operating environment similar to that of enterprise systems. Holo3 learns operations similar to those in actual work within this learning environment.
In terms of performance, the open-source version, Holo3-35B-A3B, achieved 77.8% in the international standard benchmark '
The benchmark includes 486 challenges requiring users to complete multiple steps in sequence, and is evaluated across four areas: e-commerce, business software, collaboration, and multi-app integration. This unique benchmark, named 'H Corporate Benchmark,' includes everything from relatively short operations completed within a single app to longer workflows spanning multiple applications.
According to H Company, the free tier allows users to try the open-source version of Holo3-35B-A3B via API, with a rate limit of 10 requests per minute. The higher-end model, Holo3-122B-A10B, is only available in the paid tier.
Related Posts:
in AI, Posted by log1b_ok