Meditron
EPFL's open medical LLMs, now with a fully open MeditronFO line on Apertus, OLMo and EuroLLM
Overview
Meditron is a family of open medical language models from EPFL. The 2023 Meditron-7B and 70B continued Llama 2's pretraining on a documented medical corpus, and Meditron-3 (2024) built on Llama 3.1. In 2026 EPFL's LiGHT lab released MeditronFO, Apache-licensed medical versions of the fully open Apertus, OLMo-2 and EuroLLM models with their training data and code.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 12 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Meditron
- Meditron-7B/70B (Llama 2) and Meditron-3 8B/70B (Llama 3.1) weights, after accepting the base-model licence
- Apache 2.0 MeditronFO models on Apertus, OLMo-2 and EuroLLM
- The fully-open-meditron training dataset (533,888 examples, Apache 2.0)
- Scripts to rebuild the 2023 GAP-Replay pretraining corpus
- Published medical benchmark evaluations you can re-run
- Models that run inside a hospital network with no inference traffic leaving it
Judge Meditron on transparency rather than leaderboard position. The 2023 corpus is documented, and the 2026 MeditronFO line goes further, releasing training data, code and weights under Apache 2.0, which is what you need when a reviewer asks how a clinical model was built. The older Meditron and Meditron-3 weights are Llama-licensed and gated.
It is research software: no support, no hosted endpoint, no clearance, and real engineering between the download and anything a clinician should use. Strong for institutions that must keep patient data on-site or reproduce their own model; wrong for a team that wants to buy something finished.
What we found
Meditron began in November 2023 as two EPFL models, Meditron-7B and Meditron-70B, made by continuing the pretraining of Meta's Llama 2 on a 48.1-billion-token medical corpus. The work was led by Antoine Bosselut's NLP lab with Martin Jaggi's machine-learning lab and Mary-Anne Hartley's global-health lab, LiGHT. The corpus, called GAP-Replay, mixes 46,000 clinical practice guidelines, 16.1 million PubMed abstracts, full text from 5 million open papers and a slice of general web text. About 36,000 of the guidelines are published as a ready-made dataset; the rest of the corpus has to be rebuilt with the team's download and scraping scripts.
Those first weights are open but not unrestricted. They carry the Llama 2 Community License, Hugging Face asks you to accept terms before downloading, and the original GitHub repository has not changed since April 2024. A Llama 3.1 generation, Meditron-3, followed in July 2024 under the same kind of inherited licence, with smaller variants on Gemma 2, Qwen2.5 and Phi-4 in early 2025.
The 2026 release is the one to watch. In May the LiGHT lab published MeditronFO, a recipe for turning a fully open base model into a medical one, and applied it to Apertus, OLMo-2 and EuroLLM. Training data, code and weights for those versions are released under Apache 2.0. EPFL calls it the first fully open medical framework, and its strongest model, Apertus-70B-MeditronFO, gained 6.6 points over its base on the team's medical benchmarks. The training answers were written by an open-weight model, gpt-oss-120b, with prompts co-authored by clinicians.
None of it is a product. There is no hosted service, no support and no regulatory clearance, and the original README advises against medical use without extensive further testing, including randomised controlled trials. EPFL says clinical trials at sites from Switzerland to Tanzania are being prepared. Meditron suits a hospital or research group that must keep patient data in-house and show how its model was built.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
About this listing
This entry was compiled from publicly available data including Meditron's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Meditron unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Meditron directly →
Spotted something out of date? Suggest an update →
Alternatives to Meditron
More in Healthcare
