H Company released two Holo4 models on Hugging Face this week—the 27B and 35B-A3B versions—designed specifically to view screenshots and click, type, and write code on a user's behalf. What we find notable is that the foundation is Alibaba's Qwen, and the entire model stack is fully open-source. This means that "AI operating computers"—a direction previously accessible only to big players like Anthropic and OpenAI—is now being captured by the open-source camp.
What this is
Holo4 is a vision-language model (VLM, AI that can understand images and screens). Paired with the hai-agents framework, its workflow is: screenshot → send to Holo4 for judgment → execute its commands (click, type, run code, call tools). The two versions are based on Qwen 3.8 dense architecture and Qwen 3.5 Mixture of Experts (MoE, distributing tasks across sub-models to save compute) architecture respectively. H Company, the publisher, is a European AI company founded by former DeepMind researchers.
Industry view
Supporters argue that open-source plus the Qwen foundation means any developer can locally deploy an "AI that uses your computer for you" agent, at costs far below calling Anthropic's Computer Use API. Enterprise on-premise deployment also means data doesn't need to be uploaded to American big tech.
But we see at least two risks. First, the gap between benchmark scores and real-world usability is large. Previous "computer-operating" models impressed in demos but exhibited persistently high error rates when deployed into enterprise workflows—mislocated cursors and misidentified buttons were the norm. Holo4 has no large-scale production data yet.
Second, the security boundary. An AI capable of "clicking screens, typing, and invoking code" holds the highest computer permissions; once compromised by prompt injection (hiding malicious instructions in web pages or documents to trick the AI into executing them), the consequences are far more severe than with a chatbot.
Impact on regular people
For enterprise IT: data-transfer workflows between internal systems—ERP, CRM, reporting—are likely to be transformed first. But handing employee permissions over to an open-source AI model places audit responsibility on the enterprise itself—don't roll it into production based on demo videos alone.
For individual workers: repetitive "open system A → screenshot → fill into system B" workflows will likely be replaced. New job opportunities lie in debugging and supervising the AI itself—evolving from "prompt engineer" toward "AI process supervisor."
For the consumer market: this won't reach consumers directly within 1-2 years; it's more likely to be embedded in SaaS (Software as a Service) products first. But as local deployment costs continue to fall, running an "AI assistant" on a personal computer will become increasingly realistic.