Early detection of frailty helps identify those at risk of adverse health outcomes, enabling interventions to reverse decline. Frailty is associated with changes in heart rate variability (HRV), and machine-learning-based assessment of HRV from the 6-minute walk test (6MWT) has proven effective for frailty assessment. Tabular Foundation Models (TFN), such as TabPFN, use pre-trained models to achieve accurate classification on small datasets while reducing the risk of overfitting. Therefore, the main goal of this paper is to assess whether TFNs can improve frailty detection using HRV compared to traditional machine learning models. In this study, we used the Physionet database titled "Wearable-based signals during physical exercises from patients with frailty after open-heart surgery." Single-lead ECGs were obtained from patients during the 6MWT. Frailty was evaluated using the Edmonton Frail Scale, with a score of 5 or less indicating non-frailty and scores above 5 indicating frailty. After QRS detection, 36 HRV features were extracted. These features include statistical measures, frequency-based metrics, traditional Poincare features, geometric Poincare attributes, and heart rate asymmetry indices. Several machine learning classifiers were employed for frailty assessment, including three TFNs, random forest, XGBoost, and CatBoost. To ensure robust evaluation and facilitate an objective comparison of model performance, 10 repeats of 5-fold cross-validation with all preprocessing inside the training folds was implemented. ECG data from 67 individuals (29 non-frail, 38 frail) were analyzed. The best result was CatBoost on HRV, AUC 0.626 (95% CI 0.511-0.778, permutation p = 0.083). All three TFMs ranked below CatBoost, which outperformed TabPFN in a paired test (p = 0.035). This cohort supports 80% power only for a true AUC near 0.70, so the finding is an inconclusive null rather than evidence against an HRV-frailty association. Pre-trained tabular models did not compensate for a small, range-restricted cohort.