Spatio-temporal wind fields in mountainous terrain are complex due to strong flow modifications when wind interacts with topography. Machine-learning (ML)-based wind downscaling models offer a computationally efficient alternative to high-resolution atmospheric model simulations and to dynamical downscaling, but their transferability beyond the training regions and conditions in which they were developed remains largely untested. We present an enhanced version of DEVINE, a deep-learning-based downscaling model for near-surface terrain-forced wind flow, being trained on atmospheric model ARPS simulations over Gaussian Random Field synthetic topographies with a wide range of slopes and feature shapes. Incorporating wind direction-dependent input features enables rotational data augmentation and facilitates preprocessing. Using enhanced DEVINE, we systematically assess its performance across previously unseen terrain (synthetic, alpine, coastal mountain environments), wind regimes (terrain-forced flow, sea-breeze circulation, downslope windstorm), and driving model data, using both high-resolution atmospheric wind and ground station data as reference.</p> <p>We find that DEVINE generalizes well to unseen mountain regions when flow is primarily terrain-forced. Its spatial wind speed mean absolute error (MAE) and correlation in the alpine and non-storm coastal environments are lower and higher, respectively, than the station-based results reported by previous ML wind downscaling studies. DEVINE’s own station-based performance remains broadly comparable to, or slightly improves on, these studies’ station-based results (e.g., MAE of 1.13 m s<sup>−1</sup>), despite simultaneously transferring across new topography, wind regime, and driving data. Performance degrades under complex atmosphere-terrain interactions such as a downslope windstorm (e.g., MAE of 5.6 m s<sup>−1</sup>); under non-storm conditions, however, DEVINE outperforms the dynamical model WRF-LES at stations for MAE (1.13 m s<sup>−1</sup> versus 1.23 m s<sup>−1</sup>) and matches it for normalized root mean square error (13 %). We further show that station-based and spatial-pattern evaluations can diverge substantially, including in the sign of bias, indicating that the evaluation basis itself shapes apparent performance, and that spatial evaluation may better separate genuine model limitations from local representativity effects.</p> <p>Our findings clarify strengths and limitations of the enhanced DEVINE wind downscaling model under out-of-sample conditions, and more broadly underscore the value of systematically evaluating ML wind downscaling methods across scale and under unseen conditions.