UC Berkeley launches CUA-Lite, open platform for computer-use agents unifying environments, data and training

A UC Berkeley research team has released CUA-Lite, an open platform for Computer-Use Agents (CUAs) that focuses on infrastructure rather than the model. The core argument: training and evaluating a CUA requires four components — agents, environments, traces, and a framework for evaluation and training — and today all of them are scattered across separate repositories with incompatible interfaces. CUA-Lite centralizes them behind a single action space, a single data schema and a single command that works on desktop, browser and mobile.

The most concrete contribution is Lite.OSWorld. The original OSWorld provides a faithful Ubuntu environment but ships as a full QEMU/KVM virtual machine for every task, demanding nested virtualization that most managed infrastructure does not expose. CUA-Lite reproduces the same task suite and evaluators on GNOME inside a simple Docker container. The natural concern is fidelity, and the team addresses it directly: across 13 models, scores on Lite.OSWorld match those on the OSWorld VM, so a training signal or score measured in the container transfers back to the real benchmark. That same foundation now hosts a family of sandboxes: Lite.ScaleCUA, Lite.CUAGym and Lite.CUAWorld, the last expanding to roughly 40 applications including Blender, QGIS and VS Code. The platform claims 30 thousand-plus verifiable tasks in total. Lite.OSWorld runs at 0.9 GB versus 4.1 GB for the VM, roughly 4.6 times more parallel desktops on the same hardware.

The second layer is LiteSample, a single supervised-learning schema shared across every environment, agent and task type, packaged as simple Parquet plus images. More than ten existing CUA datasets have been processed into it and released freely on Hugging Face, among them Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey and Multimodal-Mind2Web. Alongside these corpora sit fresh rollout datasets generated by running an advanced teacher model through the sandboxes for distillation to smaller models. Because model families expect different scaffolding, the framework ships with a per-model adapter that wraps the unified LiteSample into each model's dedicated training format, including history collapsing so that multiple steps share a single forward pass.

Agents and environments meet in lite.gym: screenshots up, actions down, with a single action space per platform. More than ten built-in agents — GPT, Claude, Gemini, Qwen3-VL, UI-TARS, Fara-7B, MAI-UI and others — and more than 15 integrated benchmarks covering grounding (ScreenSpot-Pro, OSWorld-G), desktop (OSWorld, OSWorld-2, WindowsAgentArena, CUABench), browser (WebArena, VisualWebArena, MiniWoB, WebVoyager, Online-Mind2Web, WebGym) and mobile (AndroidWorld, AndroidLab, MobileWorld, MobileGym). Swapping --model-id and --env-id in scripts/rollout.py is the entire interface. The same loop serves training: for SFT, the README documents fine-tuning Qwen3-VL-2B-Instruct on Lite.ScaleCUA desktop trajectories, which lifted average episodic return from 0.138 to 0.237 on the 332-task lite.osworld evaluation split, a single configuration reported on two GPU, not an independently reproduced result. For RL, environment-scored rollouts drive GRPO updates atop Slime, with a working MobileGym example covering 416 mobile tasks across 28 applications.

The stack installs via uv sync --all-extras on Python 3.12, and the lightweight sandboxes run on any Docker host without /dev/kvm, so cloud instances, CI runners and nested containers all work. For researchers and engineers building computer agents, the implication is that they can run experiments at scale on standard infrastructure without fighting nested virtualization, with assurance that the training signal remains valid on the original benchmark.