Xavier Theimer-Lienhard sits in a quiet lab in Lausanne, surrounded by screens flickering with medical texts and code—each line a step toward making AI in healthcare something doctors can truly trust. At EPFL’s Laboratory for Intelligent Global Health & Humanitarian Response Technologies (LiGHT), he and his team have launched MeditronFO, the world’s first fully open framework for building clinical large language models—a breakthrough that opens every layer of medical AI development to public scrutiny. Unlike most AI systems used in hospitals today, which operate as proprietary black boxes, MeditronFO releases not just the final model, but the full pipeline: training data, code, data-processing methods, and evaluation protocols, all publicly accessible. “We would never trust a clinician whose training can’t be verified, and the same standard should apply to AI in health care,” says Theimer-Lienhard, a Ph.D. student leading the project. That principle is now embedded in code and collaboration.
The stakes could not be higher. As AI increasingly supports diagnosis and treatment decisions in emergency rooms and clinics, the lack of transparency in commercial models has raised alarms among clinicians and regulators. MeditronFO answers that concern by being built from the ground up with clinician involvement—not as end users, but as co-creators. Through MOOVE (Massive Open Online Validation and Evaluations), doctors across Switzerland and beyond are actively auditing training materials, validating outputs, and flagging safety risks. The framework combines over 46,000 clinical practice guidelines with synthetic data derived from real-world patient cases and medical exams, all reviewed by medical professionals. Every dataset, every preprocessing step, every training decision is documented and open.
The results speak for themselves. Every MeditronFO-enhanced model outperformed its base version. The strongest, Apertus-70B-MeditronFO—built on Switzerland’s national AI model developed by EPFL and ETH Zurich—saw a 6.6 percentage point jump in performance on standardized medical exams. Published on arXiv in 2026, the framework is already being tested in real-world clinical workflows, with hospitals exploring how to adapt it for local use. But the impact goes beyond performance. “Our findings show that competitive medical AI models can be built through the active involvement of clinicians and communities,” says Professor Mary-Anne Hartley, a physician and director of LiGHT. “This creates a pathway for health systems to retain ownership of these technologies.” In an era of corporate-controlled AI, MeditronFO proves that another way is not only possible—it’s already working.