Somebody showed GPT-6 Luna photos of damaged wind turbine blades and asked it what kind of damage it was looking at. Fewer than 1 in 10 right. A 2B model, tiny by today’s standards, scored the same. Then one of them got a few hours of training on a single GPU. The small one. In Liquid AI’s own test, about 3,000 public photos of turbine blades and a short fine-tune, SFT plus GRPO on one A100, took LFM3-VL-2B from under 10% to 67% on roughly 1,400 photos it had never seen. The big model stayed…