Cua-S1-4B-0.2: RL-Trained Computer-Use Decision Model

Cua-S1-4B-0.2 is a second, separately trained version of Cua's small computer-use decision model, and what is new is how it learned. The model does not operate a computer by itself. Its model card describes two independently trained LoRA adapters, small add-on weight sets, for a frozen copy of Qwen's Qwen3.5-4B. One handles text input and one is multimodal, and both choose among a closed list of element and action options that the calling application supplies, including over multi-step runs in live interface environments.

The earlier Cua-S1-4B-0.1 was trained only by supervised fine-tuning on Cua's synthetic and converted task data, and its card says no reinforcement-learning stage was applied. For 0.2, Cua says it trained with RLOO, a reinforcement-learning method, on live computer-use tasks, rewarding the model when a task was completed, and calls it the first multimodal decision model trained that way. Cua's card for 0.1 reports that 0.2 scores higher on its held-out cross-dataset and live-environment evaluations, while 0.1 stays ahead only on the same-distribution multimodal split it was trained on, and it recommends 0.2 for new work.

Those comparisons are Cua's own, and the 0.2 card refers readers to Cua's cua-bench-s1 results for the figures and method; Cua-Bench-S1 is the benchmark Cua introduced for this kind of single-decision model. The adapters are released under Apache-2.0, while Qwen's license still governs the base model. Because the model only picks from options the caller lists, Cua's documentation for the earlier checkpoint warns that a confident choice says nothing about whether that list was safe or complete.