Missing values (MVs) remain a significant barrier to reliable proteomics analysis, particularly in single-cell proteomics, where small amounts of starting material and limits in detection drive Missing-Not-At-Random (MNAR) sparsity. Commonly used bulk proteomic imputation approaches typically address Missing-At-Random (MAR) MVs and improve replicate consistency at the expense of sensitivity for biological variation, whereas MNAR-specific strategies preserve group differences but compromise cross-replicate reproducibility. Existing imputation methods are commonly applied to bulk data and do not offer a generalised ‘off-the-shelf’ implementation that robustly addresses the significant sparsity observed in single-cell studies. Here, we introduce SoftHybrid, a continuous weighting framework that automatically balances MAR- and MNAR-oriented imputation across a dataset-derived model between missing rate and protein abundance model. Benchmarking across known ground truth samples (three-species mix) and real single-cell proteomics data showed that SoftHybrid outperforms existing methods at low inputs while meeting or exceeding previous state-of-the-art performance at the mini-bulk level. Preservation of proteomic patterns and enhanced replicate consistency improve testing significance and boost recovery of biological signals. SoftHybrid is implemented as an R package and freely available on GitHub.