UC Berkeley researchers have released CUA-Lite, an open-source platform for computer use agents (CUA) that integrates agents, environments, trajectories, evaluation, and training frameworks under a unified workspace, data schema, and command interface. The platform covers desktop, browser, and mobile environments. Its lightweight sandbox runs on any Docker host without requiring /dev/kvm support. The core component, Lite.OSWorld, replicates the OSWorld task suite and evaluator in standard Docker containers, reducing memory usage from 4.1GB to 0.9GB and increasing parallel instances by approximately 4.6 times while maintaining evaluation scores consistent with the original virtual machine version. The platform also includes LiteSample, a unified data schema, and has preprocessed over a dozen existing CUA datasets, published on Hugging Face. It integrates more than 10 agents based on models such as GPT, Claude, and Gemini, along with over 15 benchmarks spanning desktop, browser, and mobile. The framework supports supervised fine-tuning and reinforcement learning via a single command. For example, fine-tuning the Qwen3-VL-2B-Instruct model on Lite.ScaleCUA improved the average episode return from 0.138 to 0.237 on a 332-task evaluation set, an increase of about 71%.
Researchers at the University of California, Berkeley, have introduced CUA-Lite, an open-source platform designed for computer use agents (CUA). The platform unifies agents, environments, trajectories, and evaluation and training frameworks under a single workspace, data schema, and command interface, supporting desktop, browser, and mobile environments.
The lightweight sandbox runs on any Docker host without requiring /dev/kvm support. The core component, Lite.OSWorld, replicates the OSWorld task suite and evaluator within standard Docker containers, reducing memory usage from 4.1GB to 0.9GB, boosting parallel instance count by roughly 4.6 times, and delivering evaluation scores that match the original virtual machine-based version.
CUA-Lite also includes LiteSample, a unified data schema. The team has preprocessed over a dozen existing CUA datasets and published them on Hugging Face. The framework integrates more than 10 agents based on models like GPT, Claude, and Gemini, along with over 15 benchmarks covering desktop, browser, and mobile scenarios.
The platform supports both supervised fine-tuning and reinforcement learning through a single command. As a demonstration, the team fine-tuned the Qwen3-VL-2B-Instruct model on the Lite.ScaleCUA subset. On a 332-task evaluation set, the average episode return improved from 0.138 to 0.237, an increase of approximately 71%.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
