Holo Company launches Holo3, SOTA Computer Use model

Holo Company has launched Holo3, their latest model designed for enterprise automation. Holo3 is built for business environments and has achieved a 78.85% score on the OSWorld-Verified benchmark, surpassing previous industry standards for desktop computer use. This model operates with 10 billion active parameters within a 122 billion total parameter architecture, making it more resource-efficient than larger models like GPT 5.4 or Opus 4.6. The Holo3-35B-A3B variant’s weights are openly released under the Apache2 license, and API access is available with a free tier, widening its reach to both developers and enterprises.

Users include enterprise IT teams, automation engineers, and organizations seeking workflow automation. The model is public and can be accessed through the company's inference API and on Hugging Face. Holo3 employs an agentic learning flywheel, a continuous training approach that sharpens perception and decision-making by combining synthetic navigation data, out-of-domain augmentation, and curated reinforcement learning. This enables Holo3 to perform multi-step, multi-application business tasks, such as parsing data across emails, PDFs, and spreadsheets.

Compared to previous offerings, Holo3 requires fewer resources while outperforming larger models on complex business benchmarks. Early industry reactions note its production-readiness and low deployment costs. The Synthetic Environment Factory, used for training, simulates real enterprise scenarios, ensuring robust workflow handling. This launch positions Holo Company at the forefront of enterprise AI automation, aiming to move toward adaptive systems capable of mastering new business software autonomously.

Source