Article
Cua open-sources two S1 models at once and brings Jev in for the same test
Dongcha Beating AI News Flash: Cua, an open-source computer-use infrastructure project, has open-sourced two S1 models designed specifically for Computer Use decision-making, while also releasing Cua-Bench-S1 to evaluate models such as Jev separately.
This evaluation does not let models plan for themselves. Each question directly provides the current interface and a selected set of actions, such as “enter an email,” “check the consent box,” “click submit,” or “skip,” and the model makes only one choice. It cannot repeatedly inspect the screen, and there are no retries after failure. This tests the purest version of one question: What should be done next?
The two models released together take different approaches. Nano has only approximately 855,000 trainable parameters and was trained from scratch specifically to score candidate actions; the 4B version uses Qwen3.5-4B as its base and is then fine-tuned into the same kind of “multiple-choice model.” The former prioritizes extremely small size and speed, while the latter trades the original language model’s understanding capabilities for stronger generalization.
In addition to common operations such as form filling, login, and search, the evaluation includes safety judgment, chess, Doom, and JevBench. The latter tasks are deliberately excluded from model training to examine whether models can still make correct decisions after leaving familiar tasks.
Original link https://m.theblockbeats.info/flash/368639