H Company has released Holo1, a family of open-source Action Vision Language Models (VLMs) designed specifically for deep web UI understanding and precise localization. The models are based on the Qwen2.5-VL architecture and are fully compatible with transformers. Surfer-H, a web-native agent that interacts with browsers like a human, relies on the Holo1 family of open-weights models.

What Happened

The release of Holo1 marks a significant milestone in the development of computer-use agents and GUI automation. The Holo1 family includes two initial models: Holo1-3B and Holo1-7B, with the latter achieving 76.2% average accuracy on common UI localization benchmarks. The models have been released under the Apache 2.0 license and are available on Hugging Face.

The release of Holo1 is part of a larger effort by H Company to develop a suite of tools for computer-use agents and GUI automation. Surfer-H, which relies on the Holo1 family of models, is a modular architecture for complete web task automation that performs reading, thinking, clicking, scrolling, typing, and validating. It is designed to be flexible and modular, composed of three independent components: a Policy model that plans and drives the agent's behavior, a Localizer model that understands visual UIs for precise interactions, and a Validator model that confirms whether tasks are completed successfully.

Background and Context

H Company was founded in May 2024 by Charles Kantor, Karl Tuyls, Laurent Sifre, Julien Perolat, and Daan Wierstra. The company raised a $220 million seed round in May 2024, led by Accel with participation from Amazon, UiPath, FirstMark, Bpifrance, and Innovation Endeavors. H Company's first product release was Runner H, an agent orchestration system that entered closed beta in November 2024.

The development of Surfer-H and the Holo1 family of models is a significant step forward for computer-use agents and GUI automation. The models are trained to power the modules of Surfer-H and provide general VLM capabilities. Holo1 is fine-tuned from Qwen 2.5-VL-Instruct weights using a large-scale, diverse training mixture designed to imbue deep web understanding and precise state-to-action mapping.

Why it Matters

The release of Holo1 and Surfer-H has significant implications for the adult industry. The models can be used to automate tasks such as age verification, payment processing, and content moderation. The high accuracy of the models makes them well-suited for tasks that require precise localization and interaction with GUIs.

The use of open-weights models also allows for greater flexibility and customization. The models can be fine-tuned for specific tasks and domains, making them a valuable tool for adult industry platforms and operators. The release of Holo1 and Surfer-H marks a significant step forward in the development of computer-use agents and GUI automation.

What Comes Next

The release of Holo1 and Surfer-H is just the beginning for H Company. The company has announced plans to continue developing its suite of tools for computer-use agents and GUI automation. The next step will be to integrate the models with other platforms and services, making them more widely available to adult industry platforms and operators.

Key Facts

  • Holo1 is a family of open-source Action Vision Language Models (VLMs) designed specifically for deep web UI understanding and precise localization.
  • The models are based on the Qwen2.5-VL architecture and are fully compatible with transformers.
  • Surfer-H, a web-native agent that interacts with browsers like a human, relies on the Holo1 family of open-weights models.
  • Holo1 includes two initial models: Holo1-3B and Holo1-7B, with the latter achieving 76.2% average accuracy on common UI localization benchmarks.
  • The models have been released under the Apache 2.0 license and are available on Hugging Face.