Long-term language recovery after left-hemisphere glioma surgery varies substantially across patients. We investigated how well it can be predicted using clinical variables such as the postoperative decline in language processing, tumor grade, resection volume and location, extent of resection, and intraoperative language mapping. Beyond predicting the overall recovery, we examined which domains of language processing are most prone to persistent deficits. Fifty-nine patients with left-hemisphere gliomas completed the Russian Aphasia Test (RAT) before surgery, immediately after surgery, and at follow-up three to seventeen months after surgery. We modeled their average language score (Generalized Aphasia Quotient, GAQ) at follow-up using a cross-validated multiple linear regression and calculated the number of patients showing persistent deficits in each subtest of the RAT. The difference between GAQ scores at follow-up and before surgery was not significant at the group level but varied substantially across patients (mean -1.3%, range -34.2-9.2%). Our best-performing model predicted the follow-up GAQ scores with the mean absolute error of 3.5% (cross-validated R2 = 0.15). A greater decline in language processing immediately after surgery predicted worse recovery, whereas intraoperative language mapping predicted better recovery. Deficits in sentence repetition, verb production, verb and sentence comprehension, and object naming most often persisted at follow-up. The postoperative decline in language processing and intraoperative language mapping explain a substantial amount of variability in long-term language recovery. Verbal working memory and lexical retrieval, particularly that of verbs, are most prone to persistent deficits.