Cua公司推出了CUA-S1-FORMS，一个基于System One（快速直觉思维）理论的专用小模型，专门用于处理表单交互任务1。该模型包含70.6万参数，原始检查点文件大小仅为2.8MB1，通过评分可选择项而非逐token生成的方式进行决策1。
在性能表现上，CUA-S1-FORMS在整个决策集上达到99.7%的准确率，相比开源模型Jev的83.6%有显著提升1。具体而言，该模型在需要行动的步骤识别上实现100%准确率（Jev为96%），在已填充字段保留上也达到100%的准确率（Jev为74%）1。在响应速度方面，本地评分延迟仅需7-9毫秒，而托管版Jev的延迟为260-280毫秒（包含网络延迟）1。
该模型首次训练迭代耗时不足30分钟1。CUA-S1-FORMS在MIT许可证下开源了合成数据生成、训练、评估和Driver集成等相关工具1。
Cua has unveiled CUA-S1-FORMS, a compact model containing 706,000 parameters designed specifically for handling form interaction decisions 1. Built on System One theory—which emphasizes fast, intuitive decision-making—the model employs a scoring approach to evaluate available options rather than generating responses token-by-token, achieving 99.7% accuracy across the entire decision set compared to 83.6% for the open-source Jev alternative 1.
The model demonstrates superior performance on critical form tasks 1. It reaches 100% accuracy in identifying steps requiring action versus 96% for Jev, and maintains 100% accuracy in preserving already-filled fields compared to Jev's 74% 1. Beyond accuracy, CUA-S1-FORMS offers significant speed advantages, with local scoring latency of only 7–9 milliseconds, substantially faster than the hosted Jev variant which incurs 260–280 milliseconds including network overhead 1. The original checkpoint file is just 2.8 MB, and the initial training iteration required less than 30 minutes 1. Cua has released the synthetic data generation, training, evaluation, and driver integration components under the MIT license 1.
评论
还没有评论，欢迎留下第一条。