Background Some previous studies on different populations have yielded inconsistent findings with respect to the relationship between levels of high-density lipoprotein cholesterol (HDL-C) and future type 2 diabetes mellitus (T2DM) incidence. This study was designed to gain further insight into this relationship through a cohort study with a 25-year follow-up duration. Methods In total, 1462 individuals that were 55 years of age or older and were free of T2DM at baseline were enrolled in the present study. T2DM incidence among this study population was detected through self-reported diagnoses or the concentration of fasting plasma glucose. The data were derived from nine surveys conducted from 1992 to 2017. The correlation between HDL-C levels and the T2DM risk was assessed through Cox proportional-hazards model and proportional hazards model for the sub-distribution with time-dependent variables. Results Over the follow-up period, 120 participants were newly diagnosed with new-onset T2DM. When research participants were separated into four groups on the basis for quartiles of their levels of HDL-C measured at baseline, and incidence of diabetes declined with higher baseline HDL-C levels at 12.60, 9.70, 5.38, and 5.22 per 1000 person-years, respectively. Adjusted hazard ratios (HRs) were 0.98 (95% confidence interval [CI]: 0.62-1.55), 0.48 (95% CI: 0.27-0.85) and 0.44 (95% CI: 0.25-0.80) for individuals with HDL-C levels within the 1.15-1.39, 1.40-1.69, and >= 1.70 mmol/L ranges relative to participants with HDL-C levels < 1.15 mmol/L. Multiple sensitivity analyses similarly revealed reduced risk of diabetes incidence with increased HDL-C levels. Incorporating the levels of HDL-C into a multivariate model significantly enhanced the overall power of the predictive model (P values were 0.0296, 0.0011, respectively, for 5- and 10-year risk of diabetes). Conclusions Levels of HDL-C were independently and negatively associated with the risk of the new-onset T2DM among middle-aged and elderly Chinese.