H Company drops Holo3 to control your desktop

Getting an AI to draft a polite email is easy. Asking it to take over your mouse, open three legacy business apps, cross-check an invoice PDF against an internal budget sheet, and fire off approvals? That's usually where most agents fall apart.

H Company officially launched Holo3, a purpose-built family of computer-use models aimed directly at automating complex desktop workflows.

The flagship model, Holo3-122B-A10B, just posted a record-breaking 78.85% on OSWorld-Verified, the industry's go-to benchmark for measuring how well AI handles real desktop environments. Even better: it did that with a Mixture-of-Experts setup that only activates 10 billion parameters per forward pass.

Alongside the enterprise API for the flagship, H Company also dropped Holo3-35B-A3B on Hugging Face under an open Apache 2.0 license, while offering a free tier via their inference API.

Trained inside a synthetic software factory

Operating real software is notoriously brittle. If a pop-up shifts three pixels to the left, conventional vision-language models often lose their minds.

Built on top of Qwen 3.5, Holo3 relies on what H Company calls an Agentic Learning Flywheel. Instead of just scraping static screenshots, the team generated massive volumes of synthetic navigation data, programmatically injected unexpected out-of-domain edge cases, and funneled everything into a performance-tuned reinforcement learning pipeline.

To see if that synthetic training actually translates to corporate office work, they built a Synthetic Environment Factory. Coding agents literally code realistic web applications from scratch, generating verifiable enterprise tasks that are scored end-to-end.

They also tested Holo3 against their own H Corporate Benchmarks: 486 multi-step tasks across e-commerce, office tools, and multi-app orchestration.

In one test, the agent has to read equipment pricing from a PDF, check each employee's remaining budget in an internal system, and autonomously email tailored approval or rejection letters. That requires deep context retention across multiple app windows without dropping intent midway.

Why it matters

Computer-use agents have spent the last year trapped between two extremes: massive closed models like GPT 5.4 or Claude Opus 4.6 that cost a fortune to run continuously, and small open models that fail the second an app interface gets slightly weird.

H Company looks like it found the sweet spot. By pairing a 122B total parameter footprint with just 10B active parameters, Holo3 hits state-of-the-art benchmark numbers while running at a fraction of the compute overhead. For IT departments calculating the unit economics of replacing repetitive back-office grunt work, that cost reduction is everything.

Giving away the 35B version under Apache 2.0 is an aggressive move against proprietary vendors. Developers can take the weights, host them on their own hardware, and start building dedicated desktop copilots without vendor lock-in.

That said, caution is still warranted. Holo3 shines brightest in controlled, recognizable interfaces; H Company admits that true real-time adaptation to completely novel custom software is still a work in progress. Plus, giving an autonomous agent direct control over your desktop remains a massive security and compliance headache.

Still, Holo3 proves that you don't need a trillion-parameter black box to drive an operating system reliably. The race to build the ultimate digital white-collar intern just got a lot faster.