UC Berkeley researchers have released CUA-Lite, an open platform for computer-use agents. Its Lite.OSWorld benchmark reproduces OSWorld tasks in a plain Docker container without nested virtualization, and scores match the VM version across 13 models. The platform covers 30k+ verifiable tasks and 15+ benchmarks, with datasets on Hugging Face. It also supports SFT and RL with per-model adapters and a shared data schema. The repository ships no explicit license, so terms should be verified before commercial use.
No score is assigned. Sources and their independence are shown in the citation chain below.