MeditronFO is the first fully open framework for building medical large language models, to make AI in healthcare more transparent and accountable.
Medical large language models (LLMs) are increasingly being used in clinical settings. For example, AI is helping doctors in emergency rooms to flag diagnoses or support decisions. The problem is that most of these systems are proprietary: their training data, design choices, and decision-making processes are hidden from view, making independent review virtually impossible.
In response, researchers from EPFL's Laboratory for Intelligent Global Health & Humanitarian Response Technologies (LiGHT) in the School of Computer and Communication Sciences have released MeditronFO (Fully Open), a framework they used to "medicalize" several fully open base models, including OLMo, EuroLLM and Apertus, Switzerland's model developed by EPFL and ETH Zurich.
"Building on Meditron that was first released in 2023, MeditronFO, is a pipeline to create a medical version of any open large language model," explains Xavier Theimer-Lienhard, a PhD student leading Meditron at LiGHT. "We would never trust a clinician whose training can't be verified, and the same standard should apply to AI in healthcare. MeditronFO makes every stage of development publicly available, from the data used to train the models to the code, training procedures, and evaluation methods."
Many AI models marketed as "open" are only partially transparent. They release the trained model itself, but not the datasets, data-processing methods or training pipelines used to create it. This makes independent auditing difficult and limits the ability of clinicians, hospitals and regulators to understand how medical AI systems arrive at their recommendations.
From the ground up
MeditronFO was built with clinicians in the room, not as consumers of the end-product, but as participant contributing throughout the process, from curating training data to validating model outputs and highlighting potential safety concerns.
Through MOOVE (Massive Open Online Validation and Evaluations), clinicians participate directly in the ongoing evaluation, and improvement of models, helping ensure that development remains grounded in real-world clinical practice. This includes the auditing of training material and validated model outputs. The development process also included safeguards.
The framework combines publicly available medical datasets with clinician-reviewed synthetic data derived from medical examinations, clinical guidelines and realistic patient cases. All datasets, processing steps and training procedures are openly documented. The researchers combined a unique set of expert-curated clinical datasets drawn from more than 46,000 clinical practice guidelines.
"Our findings show that competitive medical AI models can be built through the active involvement of clinicians and communities with data and evaluations that reflect the settings where they will ultimately be used. This creates a pathway for health systems and communities to retain greater ownership of these technologies, rather than relying solely on external proprietary systems whose priorities may not always align with local needs," says Professor Mary-Anne Hartley, medical doctor and Director of LiGHT.
Real-world testing begins
Every MeditronFO model outperformed its original base model. The strongest results came from Apertus-70B-MeditronFO, which improved performance on medical exams by 6.6 percentage points over the underlying model.
"Our results have shown that fully open medical models are achievable and competitive. In medicine, where transparency is a prerequisite, and where the stakes are people's lives, this matters," says Theimer-Lienhard.
The launch of MeditronFO is an important milestone in the continuation of a broader research program. The team is preparing clinical trials in multiple sites from Switzerland to Tanzania that will evaluate how doctors use AI in real healthcare settings. These studies will examine whether clinicians follow or reject AI-generated recommendations and how those decisions affect patient care.
This multi-year trial project, called MED.USE, also aims to understand how AI can improve healthcare quality while reducing unnecessary treatments and interventions. "It is important to get real-world feedback based on patient outcomes," explains Hartley.
Why openness matters
The launch of MeditronFO reflects a broader debate over the future of AI in medicine, centered around transparency and accountability but also data sovereignty and fears of a growing dependence on proprietary AI platforms. The results demonstrate that fully open approaches can produce medical AI systems that are both transparent and highly competitive.
"The question is not whether AI will become part of healthcare, it already is. The question is what kind of AI ecosystem we want to build. We believe that transparency, scientific scrutiny, and meaningful participation from clinicians and patients should remain central. MeditronFO shows that openness and performance do not have to be competing priorities, and that there is a viable path toward medical AI that is both innovative and accountable," Hartley concludes.
The development of MeditronFO was supported by the Swiss AI Initiative (a collaborative effort between EPFL, ETH Zurich, and CSCS) which provided computing infrastructure and funding to enable large-scale model training, evaluation, and open release.
