Researchers have introduced InfiMed2, a new family of medical multimodal foundation models available in 4B and 27B parameter sizes. These models are designed with stage-aware data processing, utilizing a 55.68B-token corpus that combines clinical knowledge with biomedical visual evidence. The training pipeline includes adapting vision encoders, building broad medical knowledge, and focusing on evidence-based data mixtures. For supervised fine-tuning, InfiMed2 employs techniques like answer stability and correctness-constrained selection to generate more informative explanations. The 4B model, further optimized with reinforcement learning, achieved 66.73% mean accuracy on five benchmarks, outperforming the Qwen3.5-9B model, while the 27B model reached 73.72%, setting a new benchmark for open-weight models in this domain. AI
IMPACT Sets a new benchmark for open-weight medical multimodal models, potentially accelerating research and development in AI-driven healthcare.
RANK_REASON The item describes a new research paper detailing the development and performance of a novel AI model. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Guanghao Zhu
- Hugging Face
- InfiMed2
- Qwen3.5:9b
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
