UC Berkeley researchers have released CUA-Lite, an open platform designed to unify sandboxes, data, evaluation, and reinforcement learning for computer-use agents. The project addresses a specific infrastructure bottleneck: training and benchmarking these agents currently requires four fragmented components—agents, environments, traces, and a training framework—each scattered across separate repositories with incompatible interfaces. CUA-Lite consolidates these elements behind a single action space, one data schema, and a unified command, covering desktop, browser, and mobile environments.
The platform is deployable immediately. Installation uses uv sync --all-extras on Python 3.12. Its lightweight sandboxes run on any Docker host without requiring /dev/kvm, meaning cloud instances, CI runners, and nested containers all function correctly.
The VM tax, and how Lite.OSWorld removes it
The most concrete contribution is Lite.OSWorld. While OSWorld provides a faithful Ubuntu desktop, it ships as a full QEMU/KVM virtual machine per task. This requires nested virtualization, a feature most managed infrastructure does not expose. CUA-Lite reproduces the same task suite and evaluators on a GNOME desktop inside a plain Docker container.
The comparison between the two approaches is clear:
| Task | OSWorld | Lite.OSWorld |
|---|---|---|
| Runtime | QEMU/KVM VM | Docker container |
| Host requirement | /dev/kvm, nested virt | Any Docker host |
| Memory | 4.1 GB | 0.9 GB |
| Cold start | 29.9 s | 23.8 s |
| Parallelism | baseline | ~4.6× more instances |
| Task suite | OSWorld | Identical |
Fidelity is the obvious concern when swapping a VM for a container. The team addresses this directly: across 13 models, Lite.OSWorld scores match the OSWorld VM’s, so a score or training signal earned in the container transfers back to the real benchmark. The same base now carries a family of sandboxes: Lite.ScaleCUA, Lite.CUAGym and Lite.CUAWorld. The last expands into roughly 40 applications including Blender, QGIS and VS Code. In total the platform claims 30k+ verifiable tasks.
One schema for data, one adapter per model
CUA-Lite’s second layer is LiteSample, a single supervised-learning schema shared across every environment, agent and task type, shipped as plain parquet plus images. Ten-plus existing CUA datasets have been preprocessed into it and published free on Hugging Face, including Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey and Multimodal-Mind2Web. Alongside those corpora sit fresh rollout datasets generated by rolling a frontier teacher model through the sandboxes, for distillation into smaller students.
Because model families expect different scaffolding, the framework ships a per-model adapter that packs a unified LiteSample into each model’s own training format, including history collapsing so several steps share one forward pass.
Eval, SFT and RL behind one command
Agents and environments meet in lite.gym: screenshots up, actions down, with one action space per platform. More than 10 agents are built in GPT, Claude, Gemini, Qwen3-VL, UI-TARS, Fara-7B, MAI-UI and others. More than 15 benchmarks are integrated, spanning 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 whole interface.
The same loop serves training. For supervised fine-tuning (SFT), the README documents fine-tuning Qwen3-VL-2B-Instruct on Lite.ScaleCUA desktop trajectories. This lifts mean episode return from 0.138 to 0.237 on the 332-task lite.osworld eval split, a single reported configuration on two GPUs, not an independently reproduced result. For reinforcement learning (RL), rollouts scored in the environment drive GRPO updates on top of Slime, with a worked MobileGym example covering 416 mobile tasks across 28 apps.
Interactive explainer
&&&
Key Takeaways
- CUA-Lite unifies agents, environments, traces and training under one action space and one LiteSample schema.
- Lite.OSWorld runs OSWorld tasks VM-free in Docker at 0.9 GB versus 4.1 GB, roughly 4.6× more parallel desktops.
- Scores in the container match the OSWorld VM across 13 models, so training signal transfers to the real benchmark.
- 30k+ verifiable tasks, 15+ benchmarks, 10+ agents, and 20+ datasets published free on Hugging Face.
- Deployable on any Docker host, but the repository ships no explicit license yet — verify terms before commercial use.
What it means
For developers building computer-use agents, the friction of managing incompatible environments and data formats is removed. A researcher can now train a model on desktop tasks using a container that consumes one-tenth the memory of a standard virtual machine, then deploy that same model to mobile or web contexts without rewriting the evaluation code. The unified schema means datasets from different sources can be combined and fed into training loops with a single configuration change, speeding up the iteration cycle for teams working on GUI automation.
