UC Berkeley researchers release CUA-Lite, an open platform for computer-use agents

UC Berkeley researchers have released CUA-Lite, an open platform designed to unify the fragmented infrastructure of computer-use agents. By replacing heavy virtual machines with Lite.OSWorld, a Docker-based GNOME desktop environment, CUA-Lite reduces memory usage from 4.1 GB to 0.9 GB while maintaining benchmark fidelity. The platform integrates over 15 benchmarks, 10 agents, and 30,000 tasks under a single action space and the LiteSample data schema.

CUA-Lite platform release

• ▪The CUA-Lite repository does not ship with an explicit license as of September 2026, requiring users to verify terms before commercial use.
• ▪The CUA-Lite platform unifies computer-use agents, environments, traces, and training under a single action space, data schema, and command.
• ▪UC Berkeley researchers released CUA-Lite, an open platform for computer-use agents, in September 2026.

Lite.OSWorld VM-free architecture

• ▪The Lite.OSWorld component of CUA-Lite runs OSWorld tasks VM-free inside a plain Docker container on a GNOME desktop.
• ▪Lite.OSWorld reduces memory usage to 0.9 GB compared to 4.1 GB for the original OSWorld virtual machine, allowing roughly 4.6 times more parallel desktops.
• ▪Lite.OSWorld scores across 13 models match those of the original OSWorld virtual machine, ensuring training signals transfer back to the real benchmark.

Unified data schema

• ▪CUA-Lite introduces LiteSample, a single supervised-learning schema shared across every environment, agent, and task type, published as plain parquet and images.
• ▪Ten-plus existing computer-use agent datasets, including Aguvis, OpenCUA, and GUI-360, have been preprocessed into the LiteSample schema and published on Hugging Face.

Model adapter framework

• ▪The CUA-Lite framework includes built-in support for over 10 agents, including GPT, Claude, Gemini, Qwen3-VL, UI-TARS, and Fara-7B.
• ▪CUA-Lite provides per-model adapters that pack the unified LiteSample data into each model's specific training format, including history collapsing.

Training infrastructure

• ▪For reinforcement learning, CUA-Lite supports GRPO updates on top of Slime, featuring a worked MobileGym example covering 416 mobile tasks across 28 applications.
• ▪For supervised fine-tuning, the CUA-Lite documentation reports lifting the mean episode return of Qwen3-VL-2B-Instruct from 0.138 to 0.237 on the lite.osworld evaluation split using two GPUs.

Benchmark integration

• ▪The CUA-Lite platform features over 30,000 verifiable tasks across its family of sandboxes, which includes Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld.
• ▪CUA-Lite integrates over 15 benchmarks spanning grounding, desktop, browser, and mobile platforms, including ScreenSpot-Pro, OSWorld, WebArena, and AndroidWorld.