Connect with us

Hi, what are you looking for?

Top Stories

H Company Launches Holo3 AI Agent with 35B Parameters for Multi-App Task Execution

French startup H Company unveils Holo3, a multimodal AI model with 35 billion parameters capable of executing complex tasks across multiple applications.

French artificial intelligence startup H Company has launched Holo3, a sophisticated AI model capable of reading screens and executing tasks across multiple applications. The announcement was reported by online media outlet Gigazine on April 9 local time, highlighting that the open-source version, Holo3-35B-A3B, is now available for free on Hugging Face.

Holo3 is a large vision-language model designed to operate in web, desktop, and mobile environments. It not only reads screen content but is also equipped to perform context-appropriate actions, such as clicking buttons or filling out forms. This innovation transcends basic automation, enabling the model to handle complex workflows that span multiple applications. For instance, it can extract price information from a PDF, compare it with departmental budgets, and subsequently send an email to approve or reject a purchase. This ability to navigate between various formats—including PDFs, spreadsheets, and emails—allows it to read documents, conduct calculations, relay information, and maintain task continuity.

The open-source Holo3-35B-A3B is fine-tuned from Qwen3.5-35B-A3B and features a mixture-of-experts architecture comprising 35 billion total parameters and 3 billion active parameters for operation. Built as a multimodal AI, it processes both images and text, generating textual output based on the inputs received. The training process utilized open-source datasets, large-scale interaction data designed for AI, and human-reviewed annotated data, which enhances its ability to adapt to unfamiliar scenarios. Selection-based reinforcement learning was integrated into the training regimen, complementing these datasets.

H Company has established a Synthetic Environment Factory that automatically builds user interfaces and interaction environments closely resembling enterprise systems. This facility employs a code-generation agent to simulate work-related tasks, further refining the model’s operational capabilities.

The AI model’s performance has been quantified, with Holo3-35B-A3B achieving a score of 77.8 percent on the international benchmark OSWorld-Verified. A high-capacity variant, Holo3-122B-A10B, attained a score of 78.85 percent on the same benchmark, boasting 122 billion total parameters and 10 billion active parameters. In addition to these metrics, H Company introduced its proprietary H Corporate Benchmark, which comprises 486 tasks categorized into four distinct areas: e-commerce, business software, collaboration, and multi-app integration. This benchmark covers a spectrum of tasks, from quick operations within a single application to more intricate workflows that require seamless transitions between multiple applications.

Users interested in exploring the capabilities of Holo3-35B-A3B can access it through a free API, limited to 10 requests per minute. The more advanced Holo3-122B-A10B is available exclusively through a paid tier, reflecting the growing trend toward premium AI solutions in enterprise contexts.

The introduction of Holo3 illustrates a significant step forward in AI’s ability to integrate and automate tasks across disparate applications, setting the stage for enhanced productivity in various sectors. As organizations increasingly seek AI solutions that can streamline operations and improve efficiency, models like Holo3 may redefine workflows and task management in the digital age.

See also
Staff
Written By

The AiPressa Staff team brings you comprehensive coverage of the artificial intelligence industry, including breaking news, research developments, business trends, and policy updates. Our mission is to keep you informed about the rapidly evolving world of AI technology.

You May Also Like

Top Stories

Nvidia enters South Korea's AI market by launching 7 million Korean-language personas and the multimodal Nemotron3 Nano, aiming to establish market dominance.

Top Stories

Multiverse Computing unveils the LittleLamb AI model family on Hugging Face, reducing model size by 50% while enhancing performance for edge and mobile applications.

Top Stories

DeepSeek's V4-Pro eclipses GPT-5 and Claude in key benchmarks, achieving a Codeforces rating of 3,206 while undercutting OpenAI's costs by 89% per million tokens.

Top Stories

Hugging Face launches ML Intern, an open-source AI agent that surpasses Claude Code in scientific reasoning with a 32% GPQA score, offering $1,000 in...

Top Stories

Anonymous developer RizenML claims to have trained a 235M parameter language model on a single Nvidia RTX 5080 in 14 days, challenging traditional AI...

AI Research

Gizmo, an AI-powered study app, surpasses 13 million users and raises $22 million to enhance its adaptive learning features and global reach

Top Stories

Threat actors exploit the Marimo Python notebook vulnerability (CVE-2026-39987) to deploy NKAbuse malware via Hugging Face, launching 662 attacks in just three days.

Top Stories

Hugging Face's HoloTab Chrome extension enables AI models to mimic human behavior in web applications, enhancing automation without site-specific integrations.

© 2025 AIPressa · Part of Buzzora Media · All rights reserved. This website provides general news and educational content for informational purposes only. While we strive for accuracy, we do not guarantee the completeness or reliability of the information presented. The content should not be considered professional advice of any kind. Readers are encouraged to verify facts and consult appropriate experts when needed. We are not responsible for any loss or inconvenience resulting from the use of information on this site. Some images used on this website are generated with artificial intelligence and are illustrative in nature. They may not accurately represent the products, people, or events described in the articles.