Abstract: Boundary Representation (B-rep) serves as a fundamental schema for 3D CAD models, enabling precise and unambiguous encoding of both geometric and topological information. Nevertheless, existing B-rep-based net-works for model classification and segmentation often encounter challenges such as inadequate extraction of geometric features and excessive parameter sizes. To overcome these limitations, this paper introduces BRep-Geom, a lightweight dual-stream network designed for geometric feature enhancement in B-rep models. The proposed approach incorporates a multi-level geometric feature enhancement strategy, effectively supporting downstream tasks including classification and segmentation. The core contributions of BRepGeom include: a Multi-Dimensional Attention Encoder (MDA-Encoder), which leverages channel grouping and cross-dimensional attention mechanisms to efficiently extract global features under linear time complexity; a Lightweight Detail Enhancement Encoder (LDE-Encoder), which employs lightweight convolutions to capture fine-grained local geometric details; and a dynamic fusion mechanism that optimally integrates global features and local details. Comprehensive experiments on multiple public datasets demonstrate that BRepGeom achieves state-of-the-art performance in both classification and segmentation tasks, validating its efficiency and effectiveness. Furthermore, the network's geometric feature extraction relies solely on basic UV-grid input, offering a high-precision and cost-efficient solution suitable for industrial applications in CAD model processing.