UC Berkeley Researchers Release CUA-Lite, an Open Platform for Computer-Use Agents

The project is infrastructural rather than model-centric. Training and benchmarking a computer-use agent normally requires four pieces: agents, environments, action traces, and a framework for evaluation and training, which CUA-Lite's developers said were scattered across incompatible repositories. CUA-Lite puts all four behind a common interface; installing the stack requires Python 3.12 and a command such as uv sync --all-extras, and lightweight sandboxes run on any Docker host without /dev/kvm, supporting cloud instances, CI runners and nested containers.

The most concrete contribution is Lite.OSWorld, a reimplementation of OSWorld in a Docker container. The OSWorld benchmark typically ships as a QEMU/KVM virtual machine per task and requires nested virtualization; Lite.OSWorld instead runs the same task suite and evaluators on a GNOME desktop inside a plain container. The project reports memory usage of about 0.9 GB versus 4.1 GB, cold start time of 23.8 seconds versus 29.9 seconds, and roughly 4.6 times more parallel instances on a single host. Across 13 models, Lite.OSWorld scores matched the OSWorld VM baseline, indicating that signals from the container transfer back to the real benchmark.

Other sandboxes built on the same base include Lite.ScaleCUA, Lite.CUAGym and Lite.CUAWorld, the last covering roughly 40 applications, including Blender, QGIS and VS Code. In total the platform claims more than 30,000 verifiable tasks.

CUA-Lite standardizes data with LiteSample, a supervised learning schema stored as parquet files plus images. More than ten existing datasets, including Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey and Multimodal-Mind2Web, have been converted and released free on Hugging Face. The framework also ships per-model adapters that pack LiteSample into each model's native training format, with history collapsing so several steps can share one forward pass.

The lite.gym component connects agents to environments with a single command interface. Users swap models and environments by modifying --model-id and --env-id in scripts/rollout.py. More than 10 agents are supported out of the box, including models based on GPT, Claude, Gemini, Qwen3-VL, UI-TARS, Fara-7B and MAI-UI, and more than 15 benchmarks are integrated, spanning grounding, desktop, browser and mobile.

For training, the project documents a supervised fine-tuning run with Qwen3-VL-2B-Instruct on Lite.ScaleCUA desktop trajectories, raising the mean episode return from 0.138 to 0.237 on the 332-task Lite.OSWorld evaluation split. It describes this as a single configuration on two GPUs, not an independently reproduced result. For reinforcement learning, rollout scores from environments drive GRPO updates on Slime, and a MobileGym example covers 416 mobile tasks across 28 apps.

The repository currently ships without an explicit license, so users should verify terms before commercial use.