Google Releases MedGemma 1.5 and MedASR Open Models for Medical Imaging and Speech Recognition

January 13, 2026

Details

  • Google AI Developers announced MedGemma 1.5, an updated 4B open model from Google Research, optimized to run offline with enhanced performance on medical imaging, text, medical records, and 2D images.
  • New capabilities include high-dimensional CT and MRI processing, longitudinal chest X-ray interpretation, whole-slide histopathology, and information extraction from medical lab reports.
  • MedGemma 1.5 shows benchmark improvements: 3% higher accuracy on CT/MRI disease classification (61% vs 58%), 14% on MRI findings (65% vs 51%), 35% better anatomical localization in chest X-rays, and 18% improved lab report extraction F1 score (78% vs 60%).
  • Released alongside MedASR, an open model for medical automated speech recognition handling specialized healthcare vocabulary to transcribe audio and generate text prompts for MedGemma.
  • Models available on Vertex AI and Hugging Face; built on Gemma 3 architecture with fine-tuning on datasets like MIMIC-CXR, supporting developer fine-tuning for custom applications.
  • Real-world use includes Qmed Asia's askCPG interface for Malaysia's clinical guidelines, aiding decision support with multimodal image extensions.

Impact

Google's MedGemma 1.5 and MedASR releases accelerate AI adoption in healthcare, where the industry is integrating AI at twice the broader economy's rate, by providing open, fine-tunable models that handle multimodal inputs like CT volumes, histopathology slides, and medical speech with measurable gains such as 14% better MRI classification and under 5% word error rates in jargon-heavy transcription. This pressures rivals like Microsoft's Nuance Dragon Medical One and Nuance's DAX, which focus on speech but lack comparable open multimodal imaging at this scale, while narrowing the gap with closed systems from Siemens Healthineers' AI-Rad Companion by enabling offline deployment and lower barriers via Hugging Face. Market-wise, availability on Vertex AI supports scalable queries up to 10,000 per minute with GDPR-compliant security, potentially cutting diagnostic times by 30% as seen in Google Cloud cases, fueling the $175 billion telehealth sector and personalized medicine trajectories toward 70% routine diagnostics by 2030. Over 12-24 months, expect shifted R&D toward agentic systems integrating speech-to-vision pipelines, boosting funding into hybrid on-premise/cloud setups to address latency in real-time apps and advance predictive analytics from wearables.

Read more