Background: To calculate the percentage of tumor-infiltrating lymphocytes (TILs) in nasopharyngeal carcinoma (NPC) using deep-learning (DL) algorithms based on digital pathology images and differentiate the outcome. Methods: We recruited 435 patients with primary non-metastatic NPC and 63 patients with de novo metastatic NPC received immunotherapy. TIL_DL percentage was calculated using the convolutional neural network model, and its ability to differentiate metastasis risk and independent prognostic value were analyzed using Kaplan–Meier survival and multivariate analyses (MVA). Results: The median follow-up time of the training and validation cohorts was 69, and 76 months, respectively. Kaplan–Meier survival analysis showed that the 5-year distant metastasis-free survival (DMFS) and overall survival (OS) of patients with high TIL_DL were significantly better than those of patients with low infiltration. MVA showed that TIL_DL degree is an independent prognostic factor for DMFS (training cohort: HR=0.197, 95% CI: 0.077-0.503, p=0.001; validation cohort: HR=0.119, 95% CI: 0.028-0.503, p=0.004) and OS (training cohort: HR=0.418, 95% CI: 0.200-0.873, p=0.020; validation cohort: HR=0.158, 95% CI: 0.048-0.520, p=0.002). The concordance index (C-index) of TIL_DL was higher than that of the immunohistochemical CD3+, CD8+ T-cell, and CD20+ B-cell densities in terms of the DMFS and OS prediction accuracy. In an immunotherapy cohort of de novo metastatic NPC (n=63), MAV revealed that high TIL_DL percentage were an independent prognostic factor for PFS (HR=0.368, p= 0.008). Conclusions: TIL_DL percentage exhibited discriminative capabilities regarding the risk of metastasis and mortality in non-metastatic NPC, and has potential to be a biomarker for dmNPC received immunotherapy. This model will help select patients with a high risk of metastasis and provides a reference for improved individualized treatment.