BioMistral 7B

BioMistral-7B is a collection of open-source biomedical language models from a French academic collaboration (Avignon and Nantes universities, CNRS, Nantes University Hospital, Zenidoc). Starting from Mistral-7B-Instruct, the team further pretrained on PubMed Central Open Access text using CNRS's Jean Zay supercomputer, releasing base, DARE/TIES/SLERP-merged, and quantized variants — plus a multilingual medical benchmark (BioInstructQA). Paper accepted at ACL 2024.

It was among the best-performing open medical LLMs in its weight class at release, supports 8 European languages, and runs on a single 15GB GPU (or ~5GB quantized), making it a popular base for medical RAG pipelines.

Honest limits: the authors explicitly warn against clinical/production use without further alignment and testing — it is a research tool. Knowledge is frozen at training time, it can hallucinate medical facts, and newer medical models (Med42-v2, MedGemma) outperform it on benchmarks.