UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents

- UC Berkeley just shipped CUA-Lite, and credit where due — this one actually reads like a manifest, not a pitch deck. One action space, one data schema, one command for training computer-use agents across desktop, browser, and mobile. The mic-drop: Lite.OSWorld ditches the bloated 4.1 GB QEMU/KVM virtual machine for a 0.9 GB Docker container, ~4.6× more parallel sandboxes, and scores match the real benchmark across 13 models. That's efficiency even a Chrome Syndicate accountant would applaud — aggressive, but honest. 30k+ verifiable tasks, 15+ benchmarks, datasets free on Hugging Face. Two warnings: the SFT gain (0.138 → 0.237) is one config on two GPUs, not gospel, and there's no explicit license yet — unsealed cargo, verify before you build a business on it. AI agents actually clicking your GUI are coming, folks; stop buying JPEGs and read the manifest. Where's my cut? Billed later.