Paris-based H Company released Holo4 on 28 September, a family of AI agents built to operate software on a person’s behalf: clicking and typing on screens, writing and running code, and calling tools through MCP and APIs, choosing whichever suits the task. It comes in two sizes, a 27-billion-parameter dense model and a 35-billion-parameter mixture-of-experts model that uses about 3 billion at a time, with weights on Hugging Face and access through H’s own API.
Both are built on Qwen base models. H also released Holotron4 Nano, which applies the same training recipe to NVIDIA’s Nemotron 3 Nano Omni.
On OSWorld 2.0, a desktop-control benchmark, H reports Holo4 27B at 61.7%, against 81.8% for Anthropic’s Claude Opus 5.5; the mixture-of-experts model scores 30.9%. H’s pitch is cost rather than raw score: it says Holo4 competes with frontier models at a much lower cost per task. Unusually, it publishes every step of the benchmark runs behind its scores, so anyone can replay them.
Why it matters: capable open agents lower the cost of trying automation on ordinary business software, and published trajectories make the claims checkable. The scores are H’s own, and the post does not state a licence for the weights.