This paper introduces a novel and efficient architecture for deep stereo matching obtained through Zero-Shot Neural Architecture Search (NAS). Although accurate and capable of good generalization across datasets, state-of-the-art iterative stereo models are often too computationally expensive for low-power devices. In order to address this limitation, this work employs NAS to efficiently explore a search space of different layer types and hyperparameters, including efficient residual layers from Ghost Modules. Instead of relying on extensive training, the method evaluates candidate architectures using a combined zero-cost proxy score based on the AZ-NAS score and the number of parameters, thus promoting the selection of smaller, efficient models. Applied to RAFT-Stereo, this process yields a significantly smaller – 1.14M parameters, compared to the original 11M – and substantially faster network. The resulting architecture maintains competitive performance on various stereo benchmarks while running \(3\times \) faster, demonstrating the effectiveness of Zero-Shot NAS in optimizing deep stereo networks for resource-constrained environments.