Open medical language models in 2026 and their licences
MedGemma, Meditron, BioMistral and OpenBioLLM compared: sizes, published MedQA results and licence terms that matter before you build a product.
- Global
Key takeaways
- MedGemma comes in 4B and 27B multimodal variants and a 27B text model, built on Gemma 3, with the MedSigLIP image encoder.
- The MedGemma paper reports 86.2% on MedQA for the 27B text model, ahead of OpenBioLLM 70B at 78.2%.
- MedGemma's terms restrict use for direct diagnosis or treatment without regulatory approval.
- Meditron models inherit their base model's licence (Llama or Gemma terms); a 2026 Fully Open Meditron release publishes its corpus and recipe.
- Benchmark scores do not show safety in conversations. Test any model on your own scenarios.
Which open medical language models matter in 2026?
| Model | Maker | Sizes | Licence | Notes |
|---|---|---|---|---|
| MedGemma | 4B and 27B multimodal, 27B text | Health AI Developer Foundations terms | Strongest reported open MedQA result; image support | |
| Meditron | EPFL and partners | 7B to 70B across versions | Inherits Llama or Gemma terms | 2026 “Fully Open” version publishes data and recipe |
| BioMistral | Academic consortium | 7B | Apache-2.0 | Small, permissive, biomedical text |
| OpenBioLLM | Saama | 8B and 70B | Inherits Llama 3 terms | Used as a comparison baseline |
What do the published results show?
In the MedGemma paper, the 27B text model reports 86.2% on MedQA, described as the best open-model result at release apart from DeepSeek-R1, against 78.2% for OpenBioLLM 70B. EPFL reports that its Fully Open Meditron model was preferred over MedGemma 27B in 58.6% of its own comparisons. These are self-reported and use different test sets.
What is the difference between open weights and open source?
Open weights means you can download and run the model. Open source, in the strict sense, means an OSI-approved licence with no use restrictions. Most medical models are open weights with custom terms that restrict certain uses, so legal review is part of model choice.
How should a health team choose a model?
Shortlist on size, language coverage, licence and hardware cost. Then run your own evaluation: real user questions in every language, crisis and dosage scenarios, and rubrics written by clinicians. A smaller model fine-tuned on your domain often beats a larger general one on your tasks.
Frequently asked questions
Can we use MedGemma in a commercial product?
Commercial use is allowed under its terms if you respect the use restrictions, including not using it for direct diagnosis or treatment without the regulatory approvals that use would need. Read the current Health AI Developer Foundations terms.
Is a higher MedQA score a safer model?
No. MedQA measures exam-style knowledge. It says nothing about crisis handling, refusing to invent doses, or behaviour in your users' languages. Use it to shortlist, then test on your own scenarios.
Sources
- MedGemmaGoogle Health AI Developer Foundations · 2025-05
- MedGemma technical report (arXiv:2507.05201)arXiv · 2025-07
- Gemma-3-27B-MeditronFO model cardEPFL on Hugging Face · 2026
- BioMistral (arXiv:2402.10373)arXiv · 2024-02
Found something out of date? Write to contact@brandlabs.app and we will fix it.