Meditron is a suite of open-source large language models specialized for medicine, developed by EPFL’s LLM group. With around 2,200 stars on GitHub, the repository provides the models, training code, and evaluation setup behind Meditron-7B and Meditron-70B, which were among the strongest open medical LLMs of their generation.
General-purpose models often lack reliable clinical knowledge. Meditron was continually pretrained on a large corpus of medical guidelines, papers, and textbooks, then instruction-tuned for medical question answering and clinical reasoning. The project also introduced evaluation on benchmarks such as MedQA and MedBullets, and is honest about limits: the models are research artifacts, not certified medical devices.
Meditron’s training corpus, known as GAP, was assembled from clinical practice guidelines, medical textbooks, and PubMed literature, reflecting the project’s view that guideline-adherent knowledge matters more for medicine than raw web text. At release, Meditron-70B was reported to outperform GPT-3.5 on several medical QA benchmarks and to approach GPT-4 on some of them, which made the suite a reference point for later open medical models.
The suite also includes evaluation tooling for clinical knowledge and reasoning, and the team published detailed model cards describing intended use and limitations. Meditron’s weights have been widely fine-tuned by the community for tasks such as medical chatbots, exam preparation, and clinical text summarization, making it one of the most reused open medical model families.
Meditron is free and open source under the Apache-2.0 license, including the released model weights.
Medical AI researchers, clinical informatics teams, and developers building or evaluating healthcare language-model applications.