We introduce SMMILE, the first multimodal medical benchmark for evaluating in-context learning abilities of vision-language models.