What makes a base language model suitable for RL? Through controlled experiments, we identify key factors then leverage them to scale up mid-training.