The computer-use agent space has a fragmentation problem, and UC Berkeley just shipped an answer. CUA-Lite consolidates agents, environments, traces, and training under one action space and one data schema — a genuinely useful move when most CUA work is scattered across incompatible repos. The team’s pitch is infrastructural, not another model release, and that’s the point worth paying attention to.
The centerpiece is Lite.OSWorld, which strips away QEMU/KVM virtualization entirely. OSWorld’s faithful Ubuntu desktop normally demands nested virtualization that most managed cloud instances won’t expose. CUA-Lite runs the same task suite and evaluators on a GNOME desktop inside a plain Docker container at 0.9 GB per instance versus 4.1 GB — roughly 4.6 times more parallel desktops on the same hardware. The obvious fidelity question gets a direct answer: across 13 models, container scores match the VM’s. That’s not a trivial claim, and if it holds up, it removes the single biggest infrastructure barrier to CUA experimentation. The platform also expands to roughly 40 applications via Lite.CUAWorld, including Blender, QGIS, and VS Code, with 30,000-plus verifiable tasks.
On the data side, LiteSample provides one supervised-learning schema as plain parquet plus images. The team preprocessed ten-plus existing datasets — Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey, Multimodal-Mind2Web — and published them free on Hugging Face. Fresh rollout datasets from a frontier teacher model are included for distillation into smaller students. A per-model adapter handles the format translation, with history collapsing so multiple steps share one forward pass — a practical optimization for training efficiency that most academic releases skip.
The training loop is where CUA-Lite gets interesting for practitioners. One command swaps models and environments: —model-id and —env-id in scripts/rollout.py. For supervised fine-tuning, the README documents lifting Qwen3-VL-2B-Instruct from 0.138 to 0.237 mean episode return on a 332-task split — a single reported configuration on two GPUs, not independently reproduced, so treat that number as directional. RL training pairs environment-scored rollouts with GRPO updates on Slime, with a worked MobileGym example covering 416 tasks across 28 apps.
One caveat worth flagging: the repository ships no explicit license. That’s a real gap for any team eyeing commercial use, and it’s the kind of thing that stalls adoption faster than a missing benchmark. Still, for researchers tired of fighting infrastructure instead of training agents, CUA-Lite looks like a meaningful step toward making computer-use research feel less like an integration project and more like science.
