Holo3: Breaking the Computer Use Frontier

Hcompany has unveiled Holo3, a new generation of AI agents designed for autonomous computer use. The flagship model, Holo3-122B-A10B, achieves a state-of-the-art score of 78.85% on the OSWorld-Verified benchmark, the leading evaluation for desktop computer control. What makes this significant is its efficiency: with only 10 billion active parameters (out of 122 billion total), it delivers high performance at a fraction of the cost of large proprietary models like GPT-5.4 or Claude Opus 4.6. The model is engineered for production, trained to execute multi-step workflows in synthetic enterprise environments that mimic real business software.

The core innovation is Hcompany's 'agentic learning flywheel,' a continuous training pipeline focused on improving an agent's perception and decision-making. This involves generating synthetic navigation data, augmenting it with out-of-domain scenarios, and using curated reinforcement learning. To validate real-world readiness beyond benchmarks, Hcompany created a 'Synthetic Environment Factory' and the 'H Corporate Benchmarks'—a suite of 486 multi-step tasks across e-commerce, business software, collaboration, and complex multi-app setups. These tasks test an agent's ability to perform coordinated work, like extracting data from a PDF, cross-referencing budgets, and sending personalized emails.

Holo3-35B-A3B's weights are openly available on Hugging Face under an Apache 2.0 license, and a free tier is accessible via Hcompany's Inference API. The release marks a major step toward Hcompany's vision of the 'Autonomous Enterprise,' where AI agents can navigate and operate within virtually any digital business environment.