2026-07-23 · Zhanhui Zhou · Haoran Liu

KVM-free OS(World), at scale.

To eval or train a computer-use agent, you need real desktops — a lot of them. OSWorld gives you a faithful one: LibreOffice, Chrome, VS Code, files and windows. But it ships as a heavy virtual machine, and that's what keeps it from scaling. We propose a series of lightweight, KVM-free sandboxes with standardized interfaces, where we reproduce benchmarks (e.g. OSWorld) and generate verifiable tasks to train any agent.

OSWorld's VM tax: How we cut it with Lite.OSWorld

OSWorld's faithful desktop is a full VM per task. It needs /dev/kvm, nested virtualization, and heavy resources — which cloud instances, CI runners, and nested containers rarely provide, so it doesn't scale.

Lite.OSWorld (ours) keeps OSWorld's desktop and drops the VM. The same tasks and evaluators run on a GNOME desktop in a plain Docker container: no /dev/kvm and much cheaper, so it boots anywhere Docker runs, packing many per host.

OSWorld
Ubuntu.qcow2
QEMU · KVM
/dev/kvm
A desktop sealed in a VM — every task boots QEMU/KVM.
OSWorldLite.OSWorld
RuntimeQEMU/KVM VMDocker container
Host requirement/dev/kvm, nested virtAny Docker host
Memory4.1 GB0.9 GB
Cold start29.9 s23.8 s
Parallelismbaseline~4.6× more instances
Task suiteOSWorldIdentical
Same tasks and evaluators, a smaller footprint per desktop.

Lite.OSWorld gives the same result as the OSWorld VM — same task, and a score that tracks it within a few points. The same model runs the same task the same way in the container as in the VM, judged by the same evaluators, so a score or training signal earned in the container carries straight back to the real benchmark.

OSWorld
Lite.OSWorld
Same model, same task — switch tabs, or click a panel for the full rollout.
Success rate (%) · hover a point for the model
At scale

Beyond OSWorld: Scalable Training Sandboxes

The KVM-free container isn't just for OSWorld — it's a base for a family of scalable CUA sandboxes with verifiable tasks. The same base already runs more of them: browser and desktop tasks, real science desktops (GMAT flying spacecraft, PyMOL turning proteins). Efficient sandboxes to train and benchmark any agent, at scale.

Looping rollout trajectories — click a tile for the full rollout.

Lite.* environments ↗  ·  Rollout datasets ↗

Any agent, on any computer — without the heavy VM. Try it, or bring your own sandboxes: GitHub · Hugging Face · Email.