BioMedLM

BioMedLM is a 2.7 billion-parameter GPT-family language model built by Stanford CRFM together with MosaicML, trained only on biomedical abstracts and papers (formerly named PubMedGPT). It uses a dedicated tokenizer trained on PubMed abstracts, so biomedical terms are handled as a single token. After fine-tuning, it achieved the best performance at the time on MedQA biomedical question answering with 50.3% accuracy, and also showed competitive results on PubMedQA and BioASQ. It has been released to the community, and it states that its text generation capability is for research only and not for production use.

• 2.7 billion-parameter autoregressive language model with a GPT-2 architecture
• Trained only on PubMed abstracts and papers (The Pile)
• Biomedical-specific tokenizer trained on PubMed abstracts
• Best performance at the time, with 50.3% on MedQA
• Published results of fine-tuning on PubMedQA and BioASQ question answering

Enterprise (contact for pricing)

Stanford CRFM 연구용 모델로 커뮤니티에 공개, 생성 기능은 연구 목적 전용(실서비스 부적합)

공식 페이지 확인 2026-10-14(산출 규칙 2026-10-15)

• One of 13 AI tools in the Vertical AI field.
• Connected to 3 other tools — #1 in its field.