Kaikaku.AI releases Epicure, three ingredient models trained on recipes, chemistry, or both, using 4.14 million recipes across seven languages.

Kaikaku.AI, a London-based restaurant tech startup, published research on Epicure, which includes three nearly identical AI models differing only in training data: Cooc (trained on recipe co-occurrence), Chem (trained on flavor molecules from FlavorDB), and Core (blending both). Model weights and datasets are now available on Hugging Face.

The models process 4.14 million recipes from eleven sources in seven languages (Chinese, Russian, Vietnamese, Turkish, Indonesian, German, and English), cleaning about 200,000 raw terms into 1,790 clean ingredients. The chemistry-driven Chem model performs better across most property classifications (fruity, bitter, protein content) and separates regional cuisine groups more sharply than the other variants.

Kaikaku runs a robotic restaurant called Common Room in the Brunswick Centre and raised about $1.8 million in a pre-seed round in 2024. The company's Fusion machine can theoretically dispense 360 bowls per hour and uses ML-powered inventory management and 3D-printed food-safe components; the ingredient model could aid menu development, ingredient substitution during supply shortages, and scaling to new locations.