Overview
In September 2026, Berkeley RDI open-sourced CUA-Lite, an open platform for developing and evaluating computer-use agents, i.e. AI systems that operate desktop software, browsers and mobile apps the way a human would, a track represented by CUA-class models from OpenAI, Anthropic, Qwen and others. Historically this development chain was deeply fragmented: every environment had its own interface and runtime, every supervised dataset its own format, and every model its own action space, prompt format and rollout code, so connecting a model to an environment meant rewriting the same glue again and again. CUA-Lite unifies all three with three standardized abstractions.
Lite.Gym is the unified environment interface that folds 15+ benchmarks such as OSWorld, WebArena, AndroidWorld and WebGym into one reset/step/close loop with a shared screenshot observation and a single GUI action space of click/type on desktop and browser plus tap/swipe on mobile. Lite.Sample is the unified supervised data format; 10+ datasets such as Mind2Web and GUIOdyssey have been preprocessed and published free on Hugging Face. Each model family ships one harness that carries the same code across evaluation, SFT and RL, currently implemented for 14 model families including GPT, Claude, Qwen and Gemini.
The most practical contribution of CUA-Lite is its VM-free desktop sandbox: Docker containers that replicate the OSWorld desktop without requiring hardware virtualization such as /dev/kvm, so they run on any host with Docker, and ship 30k+ verifiable training tasks. Lite.OSWorld runs OSWorld's own tasks and evaluators unchanged. The project publishes a reproducible leaderboard across 13 agents and 325 tasks, with gpt-5.5 leading at 72.3% and Qwen3.5-27B first among open models at 46.2%.
Key Features
- Three standardized abstractions: Lite.Gym unifies environment interfaces, Lite.Sample unifies supervised data formats, and one harness per model carries the same code across evaluation, SFT and RL.
- 15+ benchmarks, plug and play: Desktop, browser and mobile benchmarks such as OSWorld, WebArena, AndroidWorld and WebGym share the same reset/step/close interface.
- VM-free desktop sandboxes: Docker containers replicate the OSWorld desktop without KVM hardware virtualization, shipping 30k+ verifiable training tasks; Lite.OSWorld runs official tasks unchanged.
- 10+ free datasets: Corpora such as Mind2Web and GUIOdyssey are preprocessed into the Lite.Sample format and published on Hugging Face together with fresh rollouts from frontier CUAs.
- Harnesses for 14 model families: GPT, Claude, Qwen, Gemini and others plug in directly, and the prompts a model sees during SFT match exactly what it saw in live rollouts.
- One-command eval and training: rollout.py pairs any agent with any benchmark via --model-id and --env-id, and the same harness supports RL training such as GRPO.
Use Cases
- Head-to-head CUA evaluation: compare GPT, Claude and Qwen models on OSWorld, WebArena and other benchmarks under one interface
- Low-cost data production: KVM-free Docker sandboxes with 30k+ tasks let individual developers batch rollout trajectories
- Fine-tuning open models: download Hugging Face corpora and convert them into a target model's SFT format with the unified adapter
- Reproducing the public leaderboard: the official 13-agent, 325-task ranking can be rerun as a basis for procurement or model selection
- Extending with your own environments: add new environments or datasets behind the unified interface so every agent in the community can evaluate and train on them
Pros
- Genuinely open: permissive license, datasets free on Hugging Face, no lock-in to any model or vendor
- VM-free sandboxes lower the barrier: no GPU or KVM hardware virtualization needed, Docker alone suffices
- One codebase, three uses: evaluation, SFT and RL share the same harness instead of reimplementing per stage
- Full platform coverage: 15+ benchmarks across desktop, browser and mobile fold into one abstraction
- Backed by Berkeley RDI, giving academic credibility and continued maintenance
- A reproducible public leaderboard provides a common yardstick for model comparison
Pricing
Fully open source and free; datasets are publicly downloadable on Hugging Face, and the evaluation environments and model adapters are open-source components.
Summary
CUA-Lite is an open platform for computer-use agents from Berkeley RDI. Its three abstractions, Lite.Gym, Lite.Sample and per-model harnesses, unify the environment, data and training pipeline so that any model can be plugged into any benchmark without rewriting code. The VM-free desktop sandbox with 30k+ verifiable tasks is what sets it apart from similar open-source projects, making it well suited for research institutions and developers who want to evaluate, fine-tune or reinforce CUA agents.