A new ultrasound-specific instruction dataset and benchmark show that domain-aligned training can substantially improve compact vision-language models for ultrasound understanding.