A New Milestone in Enterprise Automation
On April 1, 2026, Hugging Face announced Holo3, an AI model capable of executing complete workflows on computers. With a score of 78.85% on the OSWorld-Verified benchmark, it sets a new world record for desktop computer use agents.
What makes this achievement remarkable: Holo3 uses only 10 billion active parameters (122B total), compared to proprietary models like GPT 5.4 or Opus 4.6. The result: significantly lower inference costs for comparable or superior performance.
What Can Holo3 Actually Do?
The model was trained via an « agentic flywheel » — a continuous learning loop combining synthetic navigation, out-of-domain data augmentation, and curated reinforcement learning.
Multi-step tasks Holo3 handles:
- Extract equipment prices from a PDF document
- Cross-reference against each employee's remaining budget
- Autonomously send personalized approval or rejection emails
This workflow type, typical in enterprise environments, requires both document understanding, sustained multi-step reasoning, and coordination across separate applications.
The Synthetic Environment Factory
Holo3's major innovation is its Synthetic Environment Factory — proprietary infrastructure that automatically generates realistic enterprise environments for training.
These environments are built by coding agents that program complete websites from scenario specifications, producing verifiable tasks of varying difficulty. The model learns to navigate an « virtually infinite variety of user interfaces ».
To measure real-world performance, developers created H Corporate Benchmarks: 486 tasks across 4 categories (E-commerce, Business software, Collaboration, Multi-App setups).
Recommendations for Businesses
1. Identify candidate workflows — Focus on multi-step processes spanning multiple applications. These are prime candidates for agent-based automation.
2. Test via the free API tier — Holo3-35B-A3B is freely accessible via the inference API under Apache 2 license. Prototype without infrastructure investment.
3. Prepare validation data — Agent quality depends on representative test scenarios. Document current workflows with their typical exceptions.
Toward Adaptive Agency
Holo3 represents an intermediate step toward what its creators call Adaptive Agency: a model's ability not just to use known tools, but to learn in real-time how to navigate entirely new enterprise software.
For organizations, this evolution suggests it's time to structure documentation of internal interfaces — tomorrow's agents will need to understand your specific systems.
Sources
This article is part of the Neurolinks AI & Automation blog.
About the author: Matthieu Pesesse — IT & Media professional, 15+ years enterprise experience in AI, automation, and digital transformation.