Meditron
modelMeditron is EPFL's flagship open medical-LLM lineage, originating in the epfLLM group before the Swiss National AI Institute existed. The Meditron-3 suite (8B and 70B, built on Llama 3.1, July 2024) is co-developed with Yale (Mary-Anne Hartley's LiGHT lab) and the ICRC, with additional Gemma 2 / Qwen 2.5 / Phi 4 base-model variants (2025). It is one of the leading open medical LLMs and appears in the Stanford AI Index.
Meditron sits at the sharpest EPFL/SNAI boundary. Later versions run on the CSCS Alps supercomputer, and the 2026 MeditronFO generation (Apertus-70B-MeditronFO / -8B-MeditronFO) is a field-optimized fine-tune on top of the Apertus base model. We keep the Meditron fine-tunes in this EPFL entry and credit the Apertus base model to Swiss AI. The Meditron-CHUV variant (2025) is a hospital-specific fine-tune with CHUV (Lausanne University Hospital), a concrete clinical-adoption signal.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Meditron3-8B | 8B | Llama 3.1 base (July 2024); EPFL LiGHT + Yale + ICRC |
| Meditron3-70B | 70B | Llama 3.1 base (July 2024) |
| Meditron3 (Gemma2 / Qwen2.5 / Phi4) | — | 2025 base-model variants |
| Meditron-CHUV | — | hospital-specific fine-tune with CHUV Lausanne (2025) |
| Apertus-70B-MeditronFO | 70B | 2026 field-optimized; Apertus base (base credit → Swiss AI) |
| Apertus-8B-MeditronFO | 8B | 2026 field-optimized; Apertus base |