Paris-based AI startup H Company released Holo3 on March 31, a vision-language model built specifically for GUI agents. The model sets a new state-of-the-art on the OSWorld-Verified benchmark, outperforming proprietary systems from OpenAI and Anthropic on computer use tasks. The smaller variant runs on just 3 billion active parameters, and the open weights are available now on HuggingFace under an Apache 2.0 license.

What Happened

H Company launched two variants of Holo3, both based on the Qwen3.5 architecture and reinforced for perception and decision-making across GUI environments. The results on OSWorld-Verified, a benchmark that measures an AI model's ability to operate computer interfaces, are significant:

  • Holo3-35B-A3B: 35 billion total parameters with only 3 billion active, scoring 77.8%
  • Holo3-122B-A10B: 122 billion total parameters with 10 billion active, scoring 78.85%

Both models beat proprietary systems including GPT-5.4 and Claude Opus 4.6 on the same benchmark. The models operate across web, desktop, and mobile GUI environments, handling tasks like navigating applications, clicking buttons, filling forms, and completing multi-step workflows.

The smaller Holo3-35B-A3B model is available on HuggingFace under Apache 2.0 with open weights. H Company is also offering a free inference API for developers who want to test the model without running it locally.

Why It Matters for Creators

Computer use models are becoming the backbone of AI-powered automation. Instead of relying on custom API integrations, these models can see and interact with any application the same way a human does. That means automating workflows across tools that were never designed to talk to each other.

What makes Holo3 notable is the efficiency. The smaller model uses a mixture-of-experts architecture that keeps only 3 billion parameters active at inference time while drawing on 35 billion total. That is a fraction of what proprietary models require, which opens the door to running capable GUI agents on consumer hardware or at much lower cloud costs.

For creators working across multiple tools (design software, video editors, content management systems), a model like this could eventually handle repetitive GUI tasks without needing each application to offer an API. The open-weight release under Apache 2.0 means developers can fine-tune Holo3 for specific workflows without licensing restrictions.

This release continues the trend of open models closing the gap with proprietary ones on specialized tasks. Earlier open-source efforts like Holotron-12B pushed computer use forward, and Holo3 takes a larger step by beating the commercial leaders outright.

What to Do Next

  • Try the free API: H Company's blog post includes access to the free inference API so you can test Holo3 on your own GUI tasks
  • Download the weights: Grab Holo3-35B-A3B from HuggingFace if you want to run it locally or fine-tune for your specific use case
  • Follow development: H Company is sharing updates on X