Background and Aims: Cardiovascular risk scores guide primary prevention, but their comparative performance remains uncertain. We externally validated 16 established and contemporary cardiovascular risk-prediction models in a common primary-prevention evaluation framework. Methods: UK Biobank participants free from cardiovascular disease and cancer at baseline were included. 16 models, corresponding to 19 configurations, were implemented as published and evaluated against a harmonized incident CVD endpoint. Performance at 5 and 10 years was assessed using discrimination, calibration, Brier score, decision curve analysis, and TRIPOD+AI reporting quality. Results: Among 438,640 participants, 45,003 incident cardiovascular events occurred over a median follow-up of 13.5 years. Ten-year area under the curve ranged from 0.668 for QRISK3 to 0.734 for PREDICT, and C-index from 0.655 to 0.717. Calibration varied substantially: PROCAM, Framingham, ASSIGN, and QRISK1 overestimated risk, whereas PREVENT, PREDICT, SCORE, SCORE2, SCORE2-OP, and China-PAR generally underestimated risk. QRISK2 showed the best calibration, while PREVENT had the lowest Brier score. At higher treatment thresholds, net benefit diverged, with PREVENT and PCE performing most consistently. Composite assessment favored PREVENT, QRISK2, PREDICT, PCE, and northern China-PAR variants. Conclusions: Direct application of cardiovascular risk scores across populations can produce clinically important differences in calibration and net benefit. Model selection for primary prevention should require external validation, local recalibration, and assessment of clinical utility, rather than reliance on discrimination alone. PREVENT, QRISK2, PREDICT, PCE, and northern China-PAR variants showed the most balanced performance in this cohort.