MOSS-MED: Medical Multimodal Model Serving Medical Image Analysis

The remarkable advancements in large language models (LLMs) and large-scale visual encoders have laid a solid foundation for enhancing the capabilities of artificial intelligence (AI) in various application scenarios. In this study, we introduce MOSS-MED, a suite of models designed for biomedical vision-language understandings. MOSS-MED includes two models currently: MOSS-MED-2.5B and MOSS-MED-LLaMA, with different scale of their underlying LLMs. We employ a two-stage training pipeline that involves visual-text alignment and visual-involved instruction fine-tuning from both general domains and the medical domain. We evaluate the performance of MOSS-MED family on medical visual question answering (VQA) benchmarks. The results and cases demonstrate the impressive proficiency in medical expertise that MOSS-MED embodies.