H Company just released Holo1, an open-source family of action VLMs that makes web automation dramatically cheaper without sacrificing performance. The 7B model hits 76.2% average accuracy on UI localization benchmarks — the highest among small models — while the companion Surfer-H agent scores 92.2% on real-world web tasks. And the cost? Thirteen cents per task.
This isn’t another API wrapper. Holo1 models are based on the Qwen2.5-VL architecture and are fully compatible with Hugging Face’s transformers library. The release includes the WebClick benchmark, a dataset of 1,639 human-like UI tasks that gives researchers a concrete way to measure progress on GUI understanding. Loading the model takes a few lines of Python — point it at a screenshot, give it an instruction like “Select July 14th as the check-out date,” and it returns precise click coordinates.
Surfer-H, the agent built on top of Holo1, operates purely through the browser like a real user. It’s a modular system with three independent components: a Policy model that plans the agent’s behavior, a Localizer model for precise UI interactions, and a Validator model that confirms task completion. That separation matters — it means you can swap out individual pieces without rebuilding the whole stack. Most web agents rely on brittle custom APIs. This one doesn’t.
What’s genuinely new here isn’t just the accuracy numbers, though they’re solid. It’s the Pareto frontier claim on cost-efficiency. H Company is effectively arguing you no longer need to choose between cheap-but-dumb models and capable-but-expensive ones. Whether that holds up outside the WebVoyager benchmark remains to be seen — real enterprise workflows are messier than research evaluations — but releasing the weights openly means developers can actually stress-test that claim themselves rather than taking a vendor’s word for it. That alone is a refreshing change in an industry that increasingly treats model weights as trade secrets.
