KAIKAKU.AI Compresses All Human Cooking Into 2MB Epicure Model Trained on 4.14M Recipes

According to KAIKAKU.AI, on May 26, the London food AI startup published Epicure, a family of three AI models trained on 4.14 million multilingual recipes across seven languages. The models compress 1,790 ingredients into 2 megabytes by learning ingredient relationships rather than storing recipes.

Epicure's three variants serve different culinary purposes: Cooc learns from recipe co-occurrence patterns, Chem maps flavor chemistry by ingredient aroma compounds, and Core blends both approaches. Each model represents the same data as a mathematical coordinate system where ingredients receive a precise location based on global cooking behaviors, enabling functions like ingredient substitution and cuisine translation without hallucination risk.