Holo3 reaches 78.85% in AI computer use

Holo3, the new model from H Company, reaches a score of 78.85% on OSWorld-Verified, a test that measures whether an AI can operate a computer and complete real tasks inside applications. The company says the result sets a new high in this type of evaluation.

The model does more than answer questions. It can interpret what appears on screen, decide what to do and execute a sequence of actions: open documents, use programs, look up information and complete multi-step processes.

A large model with lower operating costs

Holo3-122B-A10B has 122,000 million parameters in total, but activates only 10,000 million during each task. Parameters are the internal values a model uses to process information and make decisions.

That architecture reduces the amount of work required for each query compared with a model that activated all its parameters at the same time. H Company says this could offer a lower cost than proprietary models such as GPT-5.4 or Opus 4.6, although the announcement does not publish a detailed price comparison.

There is also a smaller version. The weights for Holo3-35B-A3B are available on Hugging Face under the Apache 2.0 license, and the model can be tested through the company’s inference API with a free tier.

Trained to work across multiple applications

Holo3’s difference is not just its size, but how it was trained. H Company uses a continuous improvement process that combines synthetic examples, task variations and reinforcement learning, a method in which the system receives signals about which decisions work best.

The process includes:

• Generating instructions and navigation examples within digital interfaces.
• Creating variations so the model can handle unexpected situations.
• Filtering the data and selecting the most useful examples to reinforce its behavior.

To test these capabilities, H Company created synthetic enterprise environments. These are automatically generated versions of workplace tools and websites, built by coding agents and verified through scripts that check whether each task was completed correctly.

The company also designed H Corporate Benchmarks, an in-house evaluation with 486 multi-step tasks. The tests cover four areas:

• E-commerce.
• Enterprise software.
• Collaboration.
• Workflows that combine multiple applications.

A difficult task might require checking equipment prices in a PDF, comparing them with each employee’s available budget and then sending a personalized approval or rejection email to each person. Understanding text is not enough: the AI must retain context, perform calculations and coordinate actions across different programs.

What changes for you

If these results carry over to real products, AI assistants will be able to handle repetitive administrative processes instead of being limited to drafting text. For example, they could collect data from several tools, update a spreadsheet and prepare the corresponding messages without you having to guide every click.

But there is an important difference between completing tasks in controlled environments and operating freely in real systems. Errors, access permissions and sensitive information remain significant obstacles, especially when an action could send an email, modify a record or approve an expense.

Holo3 shows that autonomous computer use is advancing through both specialized training and model size. H Company’s next proposed step is for these agents to learn how to handle new enterprise software, even when they have never seen its interface before. That will show whether they can move beyond following familiar routines and truly adapt to each company’s digital work.