Hcompany/Holo4-27B

Holo4-27B is a vision‑language model designed for computer‑use tasks. Built on the dense Qwen3.8‑27B architecture, it can interpret screenshots, run code, and invoke tool calls through the hai‑agents harness. The model reports an 85.2 % score on the OSWorld benchmark, although independent verification is not provided. The model operates by receiving visual input and contextual data, then returning instructions that the harness executes as clicks, typing, or code. It supports graphical interfaces, code interaction, and tool invocation across web, desktop, and mobile environments. An example prompt asks the model to construct a detailed 3D Eiffel Tower in FreeCAD, specifying dimensions and geometry. Performance claims should be examined against the open‑source trajectory dataset released by Hcompany, as the benchmark results lack external replication. Licensing details matter: model weights are under a non‑commercial CC BY‑NC 4.0 license while the underlying Qwen3.8‑27B base is Apache 2.0. Cost per task varies by benchmark, with reported figures such as $0.08 per OSWorld task and $0.05 per AutomationBench task.