Cua open-sources two S1 models in one go, and even brings Jev in for a same-stage test.

Beating AI News Flash: Open-source computer-use infrastructure project Cua has open-sourced two S1 models specifically designed for Computer Use decision-making, and simultaneously released Cua-Bench-S1 to evaluate models like Jev in isolation.

This benchmark does not let the model plan on its own. Each question directly provides the current interface and a set of pre-selected actions, such as "fill in email," "check the agreement box," "click submit," "skip," and the model makes only one choice. It cannot repeatedly look at the screen, and there is no failure retry. What this tests is the purest single question: what should be done next.

The two models open-sourced together take two different paths. Nano has only about 855,000 trainable parameters, is trained from scratch, and is specifically designed to score candidate actions; the 4B version directly uses Qwen3.5-4B as its base and is then fine-tuned into the same "multiple-choice model." The former pursues extreme smallness and speed, while the latter trades the large model's original comprehension ability for stronger generalization.

In addition to common operations such as form filling, logging in, and searching, the benchmark also includes safety judgment, chess, Doom, and JevBench. The latter few are deliberately not trained into the model, to see whether it can still make correct decisions after leaving familiar tasks.