<p>Strawberry appearance quality recognition is essential for automated harvesting and postharvest sorting. However, in practical agricultural environments, collecting large-scale annotated datasets is labor-intensive and expensive, while real-world deployment additionally requires lightweight models capable of efficient inference on resource-constrained devices. Conventional supervised deep networks, which rely heavily on abundant labeled data and computationally intensive architectures, therefore face substantial limitations in practical applications. To address these challenges, we propose a training-free re-parameterized few-shot (FS) strawberry appearance quality recognition network (RFSAQR-Net) for data-scarce and deployment-constrained scenarios. Specifically, a structural re-parameterization strategy is adopted, which leverages multi-branch convolutions during pretraining to enrich feature representation, while converting them into a single convolution for inference without additional computational overhead. Furthermore, an adaptive subspace-based FS classifier is introduced to model each class as a low-dimensional linear subspace, enabling robust recognition by effectively capturing intra-class variability under few labeled samples. Extensive experiments are conducted on both the StrawQ-4 strawberry quality grading dataset and the StrawDISE-7 strawberry disease dataset under multiple FS settings. Experimental results demonstrate that RFSAQR-Net achieves competitive or superior performance compared with representative baselines across most FS settings, while also demonstrating strong generalization capability toward unseen strawberry disease classes. In addition, RFSAQR-Net demonstrates favorable efficiency and deployment potential for practical agricultural applications. Overall, RFSAQR-Net provides an annotation-efficient, training-free, and practically deployable solution for intelligent fruit grading and disease recognition in real-world agricultural environments.</p>