We are proud to announce Holo3, the latest evolution of our vision for the autonomous enterprise. With a score of 78.85% on the OSWorld-Verified benchmark, Holo3 establishes a new state-of-the-art in the industry on key desktop computer usage benchmarks.
Holo3 is more than just a benchmark leader. Designed for production. Built using our agent flywheel and trained to execute real-world workflows within a synthetic enterprise environment. This not only ensures that Holo3 excels in today’s business scenarios, but also establishes the foundation for a future where agents can navigate virtually any digital environment autonomously.
Best of all, Holo3 achieves this with just 10B active parameters (122B total), making it possible at a fraction of the cost of larger proprietary models such as GPT 5.4 and Opus 4.6. All models are available through the inference API. Holo3-35B-A3B weights are openly accessible on Hugging Face under the Apache2 license and freely accessible through the inference API on the free tier.
Agenttic learning flywheel
What makes Holo3 unique is its specialized training pipeline. That is, a continuous feedback loop designed to strengthen two key agent pillars: perception and decision-making.
Our training flywheel aims to teach models how to perform specific tasks from annotated examples while developing generalist skills across a virtually infinite variety of user interfaces. Here’s how to build a world-class computer usage model.
Synthetic navigation data: Generate scenario-specific navigation samples using human and generated instructions.
Extend out-of-domain: Programmatically extend scenarios and extend data to help Holo3 handle unexpected situations.
Curated reinforcement learning: All data samples are carefully selected and ingested through a pipeline that leverages advanced data filtering and reinforcement learning to maximize performance.
Beyond the raw scores, the OSWorld results serve as the ultimate proof of concept for the learning flywheel. To verify the portability to real-world business applications, we created a synthetic environment factory.
Synthetic Environment Factory & H Corporate Benchmark
This unique factory recreates the reality of Enterprise Systems and is one of the training gyms where Holo3 was trained. Our environment is automatically built using a coding agent that programs websites from scratch based on scenario specifications, generates verifiable tasks of varying difficulty, and validates them end-to-end with validation scripts.
To measure real-world readiness, we also designed the H Corporate Benchmarks, a proprietary assessment suite of 486 multi-step realistic tasks across four categories: e-commerce, business software, collaboration, and various multi-app settings.
This benchmark spans the entire spectrum of complexity, from focused single-application tasks to long-term, multi-application workflows that reflect how work is done in real life. A more difficult level (multi-app) requires agents to coordinate information across multiple systems simultaneously. For example, retrieve equipment prices from a PDF, cross-reference them with each employee’s remaining budget, and autonomously send personalized approval or rejection emails to each individual. These types of tasks require not only accurate calculations and parsing of documents, but also continuous multi-step inference throughout the application without losing state or intent.
Example of a synthetic environment created for training Holo3
In the results below, you can see that Holo3 outperforms its competitors on single application benchmarks. The performance difference between the Holo3 model and the base Qwen3.5 model reflects the influence of the agent learning flywheel. Holo3 shows the true magnitude of this specialized training by achieving a higher success rate than models with significantly more parameters while maintaining the same localization and grounding criteria.
Toward a universal agency
Holo3 is a milestone, not a destination. By building systems that can see, reason, and act within our clients’ digital platforms, we are making the autonomous enterprise a reality.
As the “synthetic environment factory” continues to evolve, agents are learning how to handle increasingly complex tasks. Now that Holo3 has mastered the interface, we’re already working on the next frontier: Adaptive Agency. At Adaptive Agency, models not only use known tools, but autonomously learn how to navigate entirely new, bespoke enterprise software in real time.
