Lovable's Model Strategy: Route to the Best Intelligence, Not the Cheapest
Lovable uses multiple frontier models and increasingly its own fine-tuned open-weight models, routing each request to whatever is most suitable. The team explicitly refuses to optimize for cost by using measurably worse models. A research team in Stockholm does post-training using reinforcement learning on the specific mistakes frontier models make inside Lovable's agent harness. The data flywheel from a million weekly projects is the core competitive moat.
