Google released the AI model ' MedGemma 1.5 4B ' specialized for medical applications and the voice recognition model ' MedASR ' on January 13, 2026. MedGemma 1.5 4B is a lightweight model that can run locally and has improved accuracy over the previous generation.

Next generation medical image interpretation with MedGemma 1.5 and medical speech to text with MedASR

https://research.google/blog/next-generation-medical-image-interpretation-with-medgemma-15-and-medical-speech-to-text-with-medasr/

MedASR Model Card | Health AI Developer Foundations | Google for Developers
https://developers.google.com/health-ai-developer-foundations/medasr/model-card

Google has released the MedGemma series of open medical models for free. MedGemma 1.5 4B is a lightweight model that can be run locally and can read and infer from medical records in text and image formats. MedASR is a speech recognition model specialized for medical phrases, and can be integrated with the MedGemma series while providing higher transcription accuracy than existing models.

The MedGemma 1.5 4B (blue) has improved text processing accuracy compared to the previous generation model, the MedGemma 1 4B (light blue).

Below is a table showing the benchmark scores of the MedGemma 1.5 4B and various models. The MedGemma 1.5 4B recorded higher scores than the general-purpose model Gemma 3 4B and the previous generation model MedGemma 1 4B, and in some tests it even outperformed the larger MedGemma 1 27B.

The previous generation of MedGemma series was designed to process not only text but also image data such as chest X-rays and pathological tissue images. MedGemma 1.5 4B has been improved to also handle 3D data such as CT images and MRIs.

In the image processing accuracy benchmark test, the MedGemma 1.5 4B (blue) also recorded a higher score than the MedGemma 1 4B (light blue).

The MedGemma series is also characterized by its ease of fine-tuning by developers,

with over 500 derivative models already released. It is expected that future development contributions will lead to the release of models optimized for various tasks, including the MedGemma 1.5 4B.

MedASR, released on the same day, is a speech recognition model specialized for medical applications. MedASR is said to be able to transcribe 'conversations about chest X-ray images' with an error rate of 5.2%. Google is highlighting MedASR's superiority over OpenAI's Whisper large-v3, which had an error rate of 12.5%. Additionally, MedASR's transcription results can also be used as prompts in the MedGemma series.

The model data for MedGemma 1.5 4B and MedASR is available at the following links:

google/medgemma-1.5-4b-it · Hugging Face
https://huggingface.co/google/medgemma-1.5-4b-it

google/medasr · Hugging Face
https://huggingface.co/google/medasr