Three new models from Kaikaku.AI distinguish between ingredient pairings based on recipes versus molecular chemistry. Trained on 4.14 million recipes and the FlavorDB database, the chemistry-focused variant outperforms recipe models in predicting nutritional values. This separation allows developers to choose between cultural taste trends and scientific flavor compatibility for food apps.