How does Holo3 achieve its benchmark score with a relatively small parameter count?

Holo3's performance is attributed to its specialized training pipeline, called the “agentic flywheel.” This system uses a continuous feedback loop with curated reinforcement learning and data from a “Synthetic Environment Factory” to specifically train the model on computer navigation and task execution, leading to higher efficiency than larger, more generalized models.

Holo3: Breaking the Computer Use Frontier

Hcompany has released Holo3, an agentic model that set a new performance record on the OSWorld-Verified benchmark for desktop computer use. With a score of 78.85%, the model demonstrates a high level of proficiency in navigating digital interfaces, suggesting a growing capability for specialized agents in enterprise automation.

The model achieves this result with a reported 10 billion active parameters, a fraction of the size of many large-scale proprietary systems. Hcompany attributes this efficiency to its “agentic flywheel” training pipeline, which uses a continuous feedback loop and a proprietary “Synthetic Environment Factory” to generate diverse, realistic business scenarios for training. This process focuses on improving the model's core abilities in perception and decision-making. The company has also made the weights for its Holo3-35B model available on Hugging Face under an Apache 2.0 license.

Holo3’s performance highlights the value of specialized training methodologies and high-quality synthetic data over simply increasing a model's parameter count. By proving its ability to handle complex, multi-application workflows—such as cross-referencing data between documents and sending customized emails—the model makes a case for the near-term viability of autonomous agents in corporate environments. This development places further emphasis on training techniques as a key differentiator in a competitive market, as Hcompany states it is now working towards agents that can learn to use entirely new software in real time.

Hcompany's results with Holo3 indicate that the path to capable autonomous agents may rely less on raw model scale and more on the sophistication of the training data and learning architecture, particularly the use of purpose-built synthetic environments.