The paper introduces a Neural Markov Random Field (NMRF) model for stereo matching, addressing the limitations of traditional hand-crafted Markov Random Field (MRF) models. The proposed NMRF model uses data-driven neural networks to design both potential functions and message passing, leveraging variational inference theory to ensure convergence and retain the graph inductive bias of MRFs. To improve efficiency and scalability, a Disparity Proposal Network (DPN) is introduced to prune the search space of disparities. The method achieves state-of-the-art performance on the KITTI 2012 and 2015 datasets, outperforming previous global methods by more than 50% in terms of the D1 metric. It also demonstrates strong cross-domain generalization and the ability to recover sharp edges. The contributions include a novel fully data-driven MRF model, a search space pruning module, and state-of-the-art results on popular benchmarks.The paper introduces a Neural Markov Random Field (NMRF) model for stereo matching, addressing the limitations of traditional hand-crafted Markov Random Field (MRF) models. The proposed NMRF model uses data-driven neural networks to design both potential functions and message passing, leveraging variational inference theory to ensure convergence and retain the graph inductive bias of MRFs. To improve efficiency and scalability, a Disparity Proposal Network (DPN) is introduced to prune the search space of disparities. The method achieves state-of-the-art performance on the KITTI 2012 and 2015 datasets, outperforming previous global methods by more than 50% in terms of the D1 metric. It also demonstrates strong cross-domain generalization and the ability to recover sharp edges. The contributions include a novel fully data-driven MRF model, a search space pruning module, and state-of-the-art results on popular benchmarks.