Recent advances in deep learning opened a new window for achieving precise free water estimation (FWE) through single-shell diffusion-weighted imaging (DWI) data. Compared with traditional approaches, those methods mitigated biases from free water contamination without relying on multi-shell data. However, current single-shell-based FWE methods still suffer from the “case-by-case” deep learning design, struggling to generalize across the diverse acquisition schemes in a more clinically useful plug-and-play fashion. In this paper, we propose a novel token-aware single-shell FWE (Ts-FWE) method towards a “one-for-all” design, so that a single model is able to train and test on different shell configurations, even for unseen data. Specifically, Ts-FWE integrates a token that encapsulates the shell configuration with a Vision Transformer (ViT) backbone architecture. Moreover, 3D patches are employed to enhance the performance compared with traditional voxel-based modeling. Both cross-validation and external validation are performed through HCP young adults, aging and MASIVar datasets. The results demonstrated that the proposed method achieved superior performance, demonstrating a promising step towards building a more generalizable deep learning scheme for DWI.