A preprint reports a compact AI model using less sampling compute and scoring better on two satellite-to-radar benchmarks, but running slower.