We’re proud to unveil Holo3, the newest evolution of our vision for the Autonomous Enterprise. With a rating of 78.85% on the OSWorld-Verified benchmark, Holo3-122B-A10B establishes a brand new cutting-edge for the industry on the leading desktop computer use benchmark.
Holo3 is greater than a benchmark leader; it’s engineered for production. Built using our agentic flywheel, it has been trained to execute real-world workflows inside synthetic enterprise environments. This not only ensures that Holo3 excels in today’s business scenarios, but establishes the inspiration for a future where our agents can autonomously navigate virtually any digital landscape.
Better of all, Holo3 achieves this with only 10B energetic parameters (122B total), so at a fraction of the associated fee of large-scale proprietary models, comparable to GPT 5.4 or Opus 4.6. All models can be found through our Inference API. Holo3-35B-A3B weights are openly accessible on Hugging Face under the Apache2 license and freely accessible through our inference API under a free tier.
The Agentic Learning Flywheel
What sets Holo3 apart is its specialized training pipeline—a continuous feedback loop designed to sharpen two core agentic pillars: perception and decision-making.
Our training flywheel is about teaching our model from annotated examples how you can execute specific tasks, all while developing generalist skills across a virtually infinite number of user interfaces. Here is how we construct world-class computer use models:
-
Synthetic Navigation Data: using human and generated instructions, we generate scenario-specific navigation examples.
-
Out-of-Domain Augmentation: we programmatically extend the scenarios and augment the information to make sure Holo3 can handle the unexpected.
-
Curated Reinforcement Learning: every data sample is fastidiously curated and ingested through a pipeline that leverages advanced data filtering and reinforcement learning to maximise performance.
Beyond the raw scores, the OSWorld results function a definitive proof-of-concept for our learning flywheel. To validate its transferability to real-world business applications we created the Synthetic Environment Factory.
The Synthetic Environment Factory & H Corporate Benchmarks
This proprietary factory reproduces the truth of enterprise systems and is one in all the training gyms Holo3 was forged in. Our environments are routinely built using coding agents that program web sites from scratch based on scenario specifications, producing verifiable tasks of various difficulty which can be validated end-to-end with verification scripts.
To measure real-world readiness, we also designed H Corporate Benchmarks, a dedicated evaluation suite of 486 multi-step realistic tasks spanning 4 categories: E-commerce, Business software, Collaboration, and various Multi-App setups.
The benchmark spans the total complexity spectrum: from focused, single-application tasks to long-horizon, multi-application workflows that mirror how work actually gets done. On the harder end of the size (Multi-Apps), tasks require the agent to coordinate information across multiple systems concurrently—for instance, retrieving equipment prices from a PDF, cross-referencing them against each worker’s remaining budget, and autonomously sending personalised approval or rejection emails to each individual. This type of task demands not only accurate calculation and document parsing, but sustained multi-step reasoning across applications without losing state or intent.
Examples of synthetic environments created for training Holo3
In our results below, we see Holo3 surpassing its competitors on single application benchmarks. The performance difference between Holo3 and the bottom Qwen3.5 models reflects the impact of our agentic learning flywheel. By achieving higher success rates than models with significantly larger parameter counts—while maintaining the identical localization and grounding standards—Holo3 illustrates the true magnitude of this specialized training.
Towards Universal Agency
Holo3 is a milestone, nevertheless it isn’t the destination. By constructing a system that may see, reason, and act inside our clients digital platform, we’re making the Autonomous Enterprise a reality.
As our “Synthetic Environment Factory” continues to evolve, our agents are learning to handle increasingly more intricate tasks. While Holo3 today masters the interface, we’re already at work on the subsequent frontier: Adaptive Agency, where our models won’t only use the tools they know but autonomously learn to navigate entirely latest, bespoke enterprise software in real-time.
