UC Berkeley Releases CUA-Lite to Unify Computer-Use Agent Development

Researchers at UC Berkeley have introduced CUA-Lite, an open-source platform that unifies agents, environments, data, and training frameworks into a single system. [1][3] The platform uses one action space, one data schema, and one command across desktop, browser, and mobile environments. [1] It installs on Python 3.12 and runs its lightweight sandboxes on any Docker host without requiring nested virtualization. [1][3]

The release includes Lite.OSWorld, which reproduces the OSWorld task suite and evaluators inside a plain Docker container instead of a QEMU/KVM virtual machine. [1] This infrastructure change reduces memory requirements per task from 4.1 GB to 0.9 GB and shortens cold start times from 29.9 seconds to 23.8 seconds. [1] The container-based setup enables approximately 4.6 times more parallel instances, and scores across 13 models match those of the original virtual machine benchmark. [1]

CUA-Lite provides a way to streamline computer-use agent evaluation by replacing resource-heavy virtual machines with lighter container infrastructure. [1] By avoiding the need for nested virtualization and `/dev/kvm`, developers can run agent training tasks on standard cloud instances and CI runners. [1] The transition allows for nearly five times the parallelism while maintaining evaluation fidelity against the original OSWorld benchmark. [1] What the sources don't address: which specific 13 models were tested to verify that container scoring matches the original virtual machine benchmark.