Holo3: Breaking the Computer Use Frontier

Hcompany has released Holo3, an AI agent that has established a new state-of-the-art score for desktop computer operation. The model achieved 78.85% on the OSWorld-Verified benchmark, a standard for evaluating an agent's ability to navigate and use computer interfaces. This result positions Holo3 as a leading system for autonomous task execution in digital environments, a critical capability for enterprise automation.

The model's performance is attributed to a specialized training method Hcompany calls an 'agentic flywheel,' which focuses on perception and decision-making. This pipeline uses synthetically generated navigation scenarios, data augmentation for handling unfamiliar interfaces, and curated reinforcement learning. Holo3 was trained in a proprietary 'Synthetic Environment Factory' and tested on a suite of 486 business tasks. With just 10 billion active parameters, the model is significantly smaller than large-scale systems like GPT-5.4, and the weights for its 35B variant are openly accessible on Hugging Face under an Apache 2.0 license.

Holo3's efficiency suggests a notable trend in AI development where sophisticated training methodologies and high-quality synthetic data can yield better performance on specialized tasks than simply increasing a model's parameter count. This approach may enable the creation of more cost-effective and targeted agents for real-world business workflows. Hcompany has indicated its future work will focus on 'Adaptive Agency,' aiming to build models that can learn to operate new and bespoke enterprise software in real-time without prior training.

Hcompany's Holo3 demonstrates that for complex agentic tasks, the sophistication of the training pipeline—specifically the use of synthetic data factories and reinforcement learning—is becoming a more critical performance driver than raw parameter count. This signals a move towards more capital-efficient, specialized models in the enterprise automation market.