Microsatellite instability (MSI) is a biomarker that helps determine eligibility for immunotherapy, but existing detection methods struggle with low tumor purity. To address this, Ziegler et al. developed MiMSI, a deep learning-based classifier using a multiple-instance learning framework, trained on NGS data with low-purity samples. MiMSI outperformed MSISensor, the current MSI detection tool, achieving a sensitivity of 0.895 and an auROC of 0.971, compared to MSISensor’s sensitivity of 0.67 and auROC of 0.907. It also reduced indeterminate classifications from 3.8% to 0.47%. In a clinical validation set of 5,037 samples across 42 cancer types, MiMSI had higher sensitivity than MSISensor (91.6% vs. 86.1%) and was significantly better for low-purity cases (<30% purity, 85.1% vs. 72.8%, p = 8.244e-07). It also improved MSI detection in prostate, endometrial, and small bowel cancers while performing comparably in colorectal and bladder cancer. Analysis of discordant cases suggested MiMSI correctly identified MSI-H tumors missed by MSISensor, while some false negatives appeared to be early-stage MMR deficiency cases without clear genomic instability. MiMSI was further validated on 45,112 tumor samples from the MSK-IMPACT sequencing panel, achieving 96% concordance with MSISensor while resolving 87% of previously indeterminate cases. It also maintained 98.6% concordance when applied to whole exome sequencing data, demonstrating robustness across different sequencing platforms.