emergentmind.comabout the medgemma family2,167 words
MedGemma 3 27B is a 27-billion parameter medical foundation model family derived from Gemma 3, available in both text-only and multimodal variants. The multimodal pathway uses a MedSigLIP vision encoder, while the text model supports a 128k-token context length and relies on post-training distillation and reinforcement learning. In EHR prescription extraction, MedGemma-27B attains 93.1% coverage and 92.2% exact-match accuracy, running locally on up to two NVIDIA L4 (24GB) GPUs. Despite strong performance in medical QA, the model exhibits prompt sensitivity and hallucination risks, necessitating schema-guided clinical deployments.written, reviewed, cited
- The text-only 27B model scores 25.7 on MedXpertQA, significantly trailing Gemma 3 27B's 15.7.
medium.comabout the medgemma family
the-decoder.com2025-07-10about the medgemma familyHN 1 pts · 0 comments696 words
Google Research and DeepMind released MedGemma, an open-source AI model suite for medical applications at this year's I/O conference. The suite includes a 4B model and a 27B version that deliver up to 10 percent higher accuracy in multimodal Q&A and 15.5 to 18.1 percent better results for X-ray classification. A companion encoder, MedSigLIP, features 400 million parameters and processes images at a 448 x 448 pixel resolution.written, reviewed, cited
- The license permits research and development but explicitly forbids direct medical diagnosis or treatment without regulatory approval.
- MedSigLIP was trained on over 33 million image-text pairs where medical data constituted only 2 percent of the total dataset.
- Reinforcement learning fine-tuning cut the error rate in electronic health record data retrieval by half.
