A new open-source AI model wants to read medical scans and answer clinical questions without making things up.
Researchers built Lingshu, a multimodal large language model trained specifically for medicine rather than adapted from a general-purpose one. The team curated a dataset pulling from medical imaging, clinical text, and general-domain sources, then synthesized captions, visual question-answer pairs, and reasoning examples from it. Lingshu was trained in multiple stages, first building medical knowledge and then layering on task-solving skills, and the researchers also tried reinforcement learning to sharpen its reasoning. To measure results, they built MedEvalKit, a standardized benchmark suite covering multimodal question answering, text-based question answering, and medical report generation, and report that Lingshu beats existing open-source medical models on most of those tasks.
Medical AI has a well-known hallucination problem, and the researchers trace it to data curation, not just model size. Models trained mostly on imaging alone tend to miss the broader clinical knowledge doctors rely on day to day. Lingshu's bet is that fixing what goes into training, and standardizing how results get measured, matters more than another architecture tweak.
Whether Lingshu holds up outside benchmarks built by its own creators is the question every medical AI claim eventually has to answer.