UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
MarkTechPost Asif Razzaq
UC Berkeley researchers released CUA-Lite, a platform for computer-use agents. It stitches sandboxes, data, eval, and training into one stack, and runs without special VM hardware.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
UC Berkeley researchers have released CUA-Lite, an open platform for computer-use agents that tries to fix a messy problem instead of chasing yet another model headline. The pitch is simple: training and benchmarking these agents needs four things — agents, environments, traces, and a way to evaluate and train them — and those pieces have been scattered across incompatible repos for too long.
CUA-Lite pulls them into one setup with one action space, one data schema, and one command. It is built to run on Python 3.12 with uv sync --all-extras, and its sandboxes work on any Docker host without /dev/kvm. That matters because managed cloud boxes, CI runners, and nested containers are usually the places where these systems get awkward fast.
The most interesting part is Lite.OSWorld. The original OSWorld benchmark uses a full QEMU/KVM virtual machine per task, which ties it to nested virtualization. CUA-Lite recreates the same task suite and evaluators on a GNOME desktop inside a plain Docker container instead. The team says the container version uses 0.9 GB of memory versus 4.1 GB for the VM, starts in 23.8 seconds instead of 29.9, and can run about 4.6 times as many instances in parallel.
They also claim the benchmark results hold up: across 13 models, Lite.OSWorld scores matched the VM version, so a result earned in the container should transfer back to the real benchmark. On top of that base, the project adds Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, with the last one covering roughly 40 applications including Blender, QGIS, and VS Code. The platform says it now includes more than 30,000 verifiable tasks.
Data and training are bundled in too. LiteSample is the shared supervised-learning schema, shipped as parquet plus images, and more than 10 existing CUA datasets have been converted into it and posted free on Hugging Face, including Aguvis, OpenCUA, ScaleCUA, GUI-360, GUIOdyssey, and Multimodal-Mind2Web. The framework also supports rollout data from a frontier teacher model for distillation into smaller students, plus adapters that fit each model family’s own training format. For SFT, the README shows Qwen3-VL-2B-Instruct going from a mean episode return of 0.138 to 0.237 on the 332-task lite.osworld split, though that is just one reported setup on two GPUs, not a reproduced benchmark.
Evaluation, SFT, and RL all run through lite.gym. The repo includes 10-plus agents, 15-plus benchmarks across desktop, browser, mobile, and grounding tasks, and a working RL example using GRPO on top of Slime. It is the kind of infrastructure release that usually gets ignored until people realize the real bottleneck was never the model alone.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of open release: boring plumbing that actually saves everyone time. The ML field loves announcing smarter agents, but most teams still lose days to glue code, VM pain, and benchmark drift. A shared stack with containerized sandboxes is less glamorous than a new model name, which is exactly why it matters.
Read more about this at: MarkTechPost