Holo4 Brings Open-Source Computer-Use Agents to the Frontier

French AI lab H has shipped Holo4, a new family of open-weight agentic models built for generalist computer use. The release ships in two sizes, a 27B dense model and a 35B-A3B mixture-of-experts variant, plus an updated Holotron4 Nano. H's pitch is that Holo4 is not another benchmark-chasing chatbot but a model that clicks and types on a screen, writes and runs its own code, and calls MCP or API tools, switching between whichever interface a real business task demands.

What happened

Holo4 is trained to operate across every interface an agent can touch: GUIs, code execution, Model Context Protocol servers, and direct API calls. H argues most agentic models are trained for a single interface, so a GUI-focused model is blind without a screen while a tool-calling model stalls in front of an application that has no API. Holo4 collapses those silos into one model that runs the same way on desktops, the web, Android, code sandboxes, and business APIs.

The models improve sharply over their Qwen base. On OSWorld 2.0, a desktop-control benchmark, Holo4 27B scores 61.7% against 81.8% for Anthropic's closed Opus 5.5, at a fraction of the cost per task, according to H's release notes. H also open-sourced every trajectory behind those scores, letting anyone replay each step at trajectories.hcompany.ai or download the dataset from Hugging Face.

Why it matters

This is a direct shot at the closed frontier labs. H is making the case that a smaller open model trained on agentic tasks rather than raw text can compete with flagship models on real workflows at dramatically lower cost. For automation teams, that means self-hosted computer-use agents with no per-token bills and no data leaving your infrastructure.

The timing also matters. Agent safety is dominating the conversation this week, with guardrail platforms shipping precisely because autonomous agents that browse, click, and run code are the systems most likely to go off-script. Open-weight agentic models put that capability, and that risk, in more hands.

Builder impact

For builders, Holo4 signals that generalist computer-use agents are becoming a commodity layer. Three takeaways stand out. First, open weights plus published trajectories lower the barrier to fine-tuning your own agent on your own workflows. Second, the multi-interface design lets one model span GUI automation and API integration in a single pipeline. Third, the cost-per-task math on OSWorld and AutomationBench suggests self-hosted agents can undercut frontier pricing on repetitive work. Watch H's Agentic Task Factory and the Holotron4 Nano release as the ecosystem matures.