Selecting market hogs' antemortem is labour-intensive and time-consuming, involving evaluations based on weight and conformation. However, most hog markets pay the producer based on pork carcass merit, which is determined at postmortem by carcass leanness percentage. The objective of this study was to predict carcass traits and composition of live animals by using multispectral thermal imaging with computer vision models. A total of 243 finishing pigs (crossbred Large White × Landrace barrows and gilts; average body weight 122 kg) were used for that purpose. Three days before slaughter, dorsal images were captured using a multispectral camera (5–15 μm wavelength range). Once pigs were slaughtered, lean depth, fat depth and leanness percentage were obtained from hot carcasses using a Destron probe. After 24 hours postmortem, chilled carcasses were fabricated into primal cuts and analyzed for leanness percentage via dual-energy X-ray absorptiometry (DEXA). Images were preprocessed, and 238 were selected based on quality and complete data. Computer vision models were trained with data augmentation techniques to predict carcass traits and classify carcasses based on lean grade index (higher lean grade indexes > 109 scores; between 57.7 to 64.2% of leanness and 80 to 105 kg of hot carcass weight). Bayesian optimization was applied to fine-tune model hyperparameters. The models showed low performance in predicting individual carcass traits and composition variables with an (RMSE of 4.93mm, an ooSR2 of 0.04) for fat depth and (RMSE of 5.77mm, and an ooSR2 of -0.14) for lean depth. The classification model moderately distinguished high and low lean-grade indexes based on DEXA lean yield (F1 score: 0.73), while Destron assessments showed a lower F1 score (0.38). Multispectral imaging technology could enable producers to market hogs based on the best grid grade. Future research should focus on increasing sample size, and integrating additional measurements, like phenotypic (e.g., body weight, sex classification, feed efficiency, and age), and genomic data, (e.g., breed type, sire, and dam lineage) and advancing from 2D to 3D imaging to enhance model accuracy and reliability.