KAIKAKU.AI published Epicure, a family of three AI models trained on 4.14 million multilingual recipes from 11 datasets across seven languages. The models do not store recipes but instead encode ingredient relationships into a 2MB coordinate table of 1,790 ingredients, each described by 300 numbers. Using a steering operator called SLERP rotation, users can mathematically navigate cooking knowledge, such as rotating a seed ingredient toward a cuisine direction to discover substitutions or flavor pairings. The three variants—Cooc (recipe co-occurrence), Chem (flavor chemistry from FlavorDB), and Core (mix of both)—address different culinary questions from the same compact file. Unlike generalist chatbots, Epicure has no hallucination risk, knowing only its 1,790 ingredients. It surpasses the previous state-of-the-art FlavorGraph (2021) by leveraging a multilingual corpus over four times larger. Practical applications include chef substitution queries, food product development, and recipe app ingredient swaps. The models are available on Hugging Face, with an interactive map at epicure.kaikaku.ai, though training code is not yet released.

Key facts

• Epicure is a family of three AI models trained on 4.14 million multilingual recipes.
• The model stores ingredient embeddings (1,790×300) in a 2MB file, not recipes.
• SLERP rotation allows mathematical navigation of cooking knowledge by cuisine.
• Three variants: Cooc (recipe co-occurrence), Chem (flavor chemistry), Core (mixed).
• Practical uses include ingredient substitution and recipe app development.