H Company's Holo3 Is the Best Open-Source Computer-Use Agent Ever Built — 78.85% on OSWorld

What Is Holo3 — and Why Does 78.85% Matter?

OSWorld is the hardest benchmark for computer-use agents. It tests whether an AI can complete real tasks inside a desktop OS — opening apps, navigating GUIs, filling forms, managing files — exactly like a human would. A score of 78.85% means Holo3 completes more than three out of four real desktop tasks autonomously. For reference, most commercial models that weren't purpose-built for this task struggle to clear 60%.

H Company is a Paris-based AI lab that has been quietly building since 2023. Holo3 is their flagship computer-use model, and it uses a mixture-of-experts (MoE) architecture: 122B total parameters, but only 35B activate per forward pass — making inference far cheaper than a dense 122B model would be.

Benchmark Comparison: Holo3 vs. Other Computer-Use Models

What Can Holo3 Actually Do?

Holo3 targets three enterprise automation categories: e-commerce operations (product listing, inventory, order management), business software workflows (CRM updates, spreadsheet processing, email triage), and cross-app coordination (extract data from one app, reformat, enter into another — no API required).

The "no API required" point is significant. Most automation tools — including Zapier, Make, or code-based agents — need structured APIs to connect software. Holo3 operates at the pixel level, seeing the screen and clicking like a human, which means it works with any software regardless of whether it offers an API.

How to Access Holo3 Today

H Company published Holo3 weights and inference API on Hugging Face. The Apache 2.0 license means you can use it in commercial products with no royalties or restrictions. Developers can run the model via the free Hugging Face inference endpoint or self-host on GPU hardware. No waitlist, no API key approval process.

The Open-Source vs. Proprietary Computer-Use Race

Six months ago, closed models had a 15-20 point lead on OSWorld. That gap is now essentially closed. Holo3 does not just match commercial computer-use models — it outperforms them on the most rigorous benchmark available, while being free and fully open.

This follows a pattern we've seen in language benchmarks (Llama, Gemma, Mistral all closing on GPT-4) repeating in a harder domain: agentic real-world task completion. H Company's edge is the Synthetic Environment Factory — a training data engine that can scale arbitrarily as long as enterprise software keeps generating new GUI patterns to learn from.

For enterprises evaluating computer-use agents, Holo3 changes the build-vs-buy calculus. A team comfortable with Hugging Face deployment can now access frontier-quality automation for free, instead of paying $30-50 per 1,000 API calls to commercial providers.

Holo3 vs. Happycapy: Which Is Right for You?

Holo3 and Happycapy serve different audiences despite overlapping use cases. Holo3 is a foundational model — a component that engineering teams integrate into larger systems. It requires GPU infrastructure, Python knowledge, and hands-on prompt engineering to use effectively in production.

Happycapy is a complete AI workspace — you log in, describe your task in plain English, and agents handle the work. No model weights, no GPU costs, no system prompts to write. For solo operators, small teams, and business professionals who need AI automation but not an engineering project, Happycapy is the faster path to productivity.

Frequently Asked Questions