Premium cocoa beans are important for fine chocolate manufacturing. However, cocoa beans may suffer from improper management practices that reduce their final quality. The objective of this study was to establish a non-invasive and high throughput grading system for cocoa bean (whole seeds) using hyperspectral imaging technique, in combination with advanced machine learning methods. Six hundred cocoa beans were collected and scanned using a HySpex Classic SWIR camera covering the spectral range from 970 to 2500 nm, with a spatial resolution of 250 μm, and a spectral sampling of 5.45 nm. Each bean was then graded using cut test methodology, the internationally recognized standard procedure in the market for cocoa trade. A maximum entropy multiclass classification model was built based on the hyperspectral cube and results from the cut test. Cocoa beans were identified into different classes: good beans, under-fermented beans, slaty beans, and other low-quality beans. For the most critical classes (good, under-fermented and slaty), a classification accuracy close to 80% was achieved without having to cut the beans open. The classification model can also distinguish other defects such as germination, over-fermentation, mold, and white beans. The proposed hyperspectral solution can significantly increase the onsite evaluation capabilities for large number of samples, potentially applicable to full batches of cocoa beans. The analysis of all beans in a batch can provide a more reliable assessment of the overall quality, compared to the results traditionally obtained from cut tests using small sample sets from batches of several tons.