Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
Decision: adopt it for research, evaluation, auditing and as a base for medical specialisation. Do not adopt it for clinical use. The publisher says so in its own words: "MeditronFO models are not medical devices and have not been tested in clinical practice. Don't use them to make decisions about a patient." (project page, ev-project-page).
The model cards say the models are "intended to be used for Medicine related tasks evaluation" and should be "assistive tools rather than definitive sources of information" (ev-card-70b). The dataset card says models trained on the corpus "should not be deployed without independent safety evaluation" (ev-data-card). The website chat describes it as "a research model, not a doctor" (ev-website).
| Use | Status | Basis |
|---|---|---|
| Research on medical LLMs, benchmarking, auditing the open pipeline | Cleared by the publisher's stated intent | cards, dataset card |
| Specialisation base for your own medical fine-tune | Cleared (open weights, corpus and training code); licence chain needs the checks under the licence item | ev-github-readme |
| Internal, non-patient drafting or education with a clinician reading every output | Possible at Deployment Ceiling T1 with your own safety evaluation; not covered by any publisher evaluation | OneHill ceiling |
| Any output that informs a decision about a patient | Not cleared: the publisher prohibits it in plain words | ev-project-page |
| Use as, or inside, a medical device or clinical decision support product | Not cleared: no device status, no clinical testing, no regulatory file; the publisher announces clinical trials as future work | ev-project-page, ev-press |
| Autonomous or patient-facing deployment | Out of scope | ev-project-page |
The maker is the LiGHT laboratory at EPFL (Switzerland), with Swiss AI Initiative funding and CSCS compute (ev-card-70b). The dossier covers the two Apertus-based checkpoints, Apertus-70B-MeditronFO (the flagship) and Apertus-8B-MeditronFO. Both are supervised fine-tunes of the Apertus-Instruct models for English medical question answering and open-ended clinical reasoning.
Known limitations, bias & failure modes
All figures below are publisher statements, and no independent evaluation exists.
- The model is behind the closest open competitor. On the five-benchmark medical average Apertus-70B-MeditronFO scores 54.56 against 63.75 for MedGemma-27B, a smaller model (
ev-project-page). - General capability falls. The general-purpose average falls for four of the six fine-tunes, mostly through IFEval (down 17.6 to 31.4 points) (
ev-project-page). - Answers are longer. Part of the HealthBench gain is length: after length adjustment the 70B scores 40.3 against 35.7 for its base (
ev-project-page). - The judges are not clinicians. Agreement on safety is low, kappa 0.311 between judge and clinicians against 0.306 between clinicians. The paper calls the judge unsuitable as a deployment-readiness signal for harmlessness and fairness (
ev-project-page,ev-paper). - Corpus quality is uneven. Two LLM judges flag 9.7 to 12.2 percent of sampled rows of the final corpus. About 63 percent of questions and all answers come from one teacher, gpt-oss-120b. Decontamination is syntactic and not semantic, and clinician review covered prompts and sample outputs and not individual items (
ev-project-page,ev-data-card). - The corpus is English only and geographically skewed. It draws mostly on North American, European and Indian exam material, partly offset by AfriMed-QA (
ev-data-card). - Multiple-choice answers are distilled from a teacher. Answers were kept only if they reached the gold label in 8 samples, and open-ended answers and vignettes had no such filter (
ev-data-card). - The card says generated content may be factually inaccurate, inconsistent or biased (
ev-card-70b).
Openness tier & components
Tier open_weights_recipe. The project page claims the first fully open medical models. The base is fully open (Apertus: open weights, data and training details per its card, ev-base-card). For this release:
| Component | State | Evidence |
|---|---|---|
| Weights | open (BF16 safetensors, ungated) | ev-hf-meta |
| Training data | partial: 533,888-row corpus released with source fields; intermediate pipeline outputs, MOOVE vignettes and some upstream inputs not released | ev-data-card, ev-github-readme |
| Training code | open: Axolotl configs, data generation, resampling, decontamination | ev-github-readme, ev-code-license |
| Evaluation | partial: harness, HealthBench and AutoMOOVE code public; the MOOVE evaluation vignettes are not public | ev-github-readme |
| Documentation | open: cards, dataset card, paper, project page | |
| Licence | partial: declared Apache-2.0, no LICENSE file with the weights | ev-hf-meta |
The project page says the 534k examples were "released with the code and outputs of every step", and the EPFL press item says "every stage of development publicly available" (ev-project-page, ev-press). The GitHub README says "The intermediate files and per-item evaluation outputs (tens of GB) are not released yet" and that the MOOVE vignettes "are not public". Its known gaps list that section 1 "cannot be rerun from a fresh clone yet" (ev-github-readme). The README is taken as authoritative, so the tier is open_weights_recipe and not fully_open.
License terms & permitted use
Apache-2.0 is declared and commercial use is permitted on that declaration. Four loose ends mean legal review is needed before relying on it.
- Both weight repositories declare
license: apache-2.0in the card metadata and link the Apache 2.0 text on apache.org from the card header (ev-card-70b,ev-card-8b). Neither repository contains a LICENSE or NOTICE file (raw LICENSE is HTTP 404), and no copyright holder is named beyond "Authors: LiGHT" (ev-hf-meta). The Apache-2.0 LICENSE in the GitHub repository covers the code and not the weights (ev-code-license). - The corpus licence is stated two ways. The arXiv paper (v2, 29 May 2026) says "We release the corpus under a research-use license" and that assets are "restricted to research purposes". The dataset card (last modified 17 Nov 2026) and project page say Apache-2.0 (
ev-paper,ev-data-card,ev-project-page). The weights were retrained after the paper (September and October 2026), so the cards may be newer, but no document says the licence was changed on purpose. - Source dataset terms differ. The dataset card lists LiveQA-Med as "Open / research use" and AfriMed-QA as CC BY-SA 4.0 (the paper says CC BY 4.0), both marked "Verify the upstream terms before redistribution". The paper lists the guidelines corpus as "Per-source (mixed); research use" and the MOOVE training split as "Research use" (
ev-data-card,ev-paper). - The Apertus base terms may carry over. The base card carries an Acceptable Use Policy with an indemnity clause in its gate prompt and advises applying a personal-data output filter every six months; Hugging Face reports the gate off. The MeditronFO cards do not mention it, and no document says whether it binds a fine-tune (
ev-base-card). - Teacher outputs add no restriction as published. Every training answer was written by gpt-oss-120b, which is Apache-2.0 with a one-sentence usage policy requiring compliance with law (
ev-teacher). - The licence grants and restricts no medical use. The research and not-for-patients statements are use statements on the cards and project page and are not licence terms. The Apache-2.0 declaration itself carries no field-of-use limit.
Follow-up entries are needed for the checkpoints with a diverging base or licence. These are OLMo-2-32B-MeditronFO (base OLMo-2-0325-32B-SFT, Apache-2.0 per its card), EuroLLM-22B-MeditronFO (base EuroLLM-22B-Instruct is not readable, HTTP 401) and EuroLLM-9B-MeditronFO (base is auto-gated), each declared Apache-2.0. The others are Gemma-3-27B-MeditronFO (Gemma Terms of Use, a different licence, and the card still carries the fully-open tag), the older Meditron3 line (Llama 3.1, Gemma, Apache-2.0 and MIT tags by base) and Gemma-3-27B-MeditronFO-Amharic (no licence tag) (ev-siblings).
Supply-chain & provenance
- Source:
EPFLiGHT/Apertus-70B-MeditronFOandEPFLiGHT/Apertus-8B-MeditronFOon Hugging Face. The organisation is named "Laboratory for Intelligent Global Health & Humanitarian Response Technologies" but is not a verified organisation (isVerifiedfalse, team plan, two members) (ev-hf-meta). The EPFL AI Center press item names the LiGHT laboratory as the releasing group, and the project page sits on an EPFL-linked site, which corroborates the identity (ev-press,ev-project-page). - Format and integrity: safetensors only (30 shards for 70B, 4 for 8B), every shard has an LFS sha256, and there is no pickle or
.binfile (ev-hf-meta). No signature, attestation or scan result is published. The checklist score is about 3 of 8. - The weights are mutable. The 70B main branch was retrained on the de-truncated corpus on 2026-09-10 and replaced with the "ICLR run" on 2026-10-06, and there are no tags. Pin the commit sha (70B
d13151ac9ff56b966dd5257d5064876685a750e3, 8Bc0f8a017b6abaed4335e68a5f6c5e0096c17115f) and verify the shard hashes (ev-hf-meta). - The lineage is well documented: base
swiss-ai/Apertus-*-Instruct-2509, 533,888-row corpus withteacher_model,source_datasetandsource_idfields, answers by gpt-oss-120b (ev-card-70b,ev-data-card,ev-teacher). - Community conversions (GGUF, MLX) exist and are not canonical; none are opened or verified (
ev-quants).
EU AI Act posture and medical-use caveats
- GPAI status: a general-purpose language model fine-tuned for a domain. The Apertus-70B base is estimated at about 6.4e24 FLOPs by the 6ND rule (70.6B parameters, 15T tokens), under the 1e25 systemic-risk line. The fine-tune used 199 GPU-hours and adds little. This is a OneHill estimate from card figures and not a publisher statement (
ev-card-70b,ev-base-card). - Open-source exemption: not established. The declared licence would normally qualify, but the missing LICENSE file and the corpus-licence disagreement leave it unsettled. Whether fine-tuning makes EPFL or the deployer a provider depends on the modification-compute rule, which is not assessed here.
- Provider documentation: the base publisher links an EU AI Act public summary and a Code of Practice document on the Apertus card (
ev-apertus-eu-docs, documented but unverified). MeditronFO publishes no provider package of its own. The open corpus with source fields and the paper help a deployer build a training-content summary. - Medical purpose is the live regulatory question, and it belongs to the adopter. The publisher states the models are not medical devices and untested in clinical practice (
ev-project-page). A product that informs clinical decisions needs a medical-device software and AI Act high-risk assessment before development. Those instruments are not assessed here, and nothing in the release is a regulatory clearance of any kind.
Benchmarks & evaluation
All results are the publisher's, run with an LLM grader (Gemma-4-31B for HealthBench) and greedy decoding for multiple choice. No independent re-run exists.
| Benchmark | Apertus-70B-Instruct | Apertus-70B-MeditronFO | MedGemma-27B (project page) |
|---|---|---|---|
| MedMCQA | 52.8 | 60.2 | |
| MedQA | 61.9 | 71.1 | |
| PubMedQA | 74.7 | 74.9 | |
| MedXpertQA | 13.0 | 17.3 | |
| HealthBench (5,000 conversations) | 39.4 | 49.3 | 54.8 |
| Five-benchmark average | 48.35 | 54.56 | 63.75 |
The 8B scores 48.26 against 42.39 for its base (ev-card-8b). On AutoMOOVE (516 clinician-written vignettes, three LLM judges) the 70B is preferred over its base in 89.7 percent of comparisons and over MedGemma-27B in 35.8 percent, losing to it under all three judges (ev-project-page).
The arXiv paper v2 reports different numbers for the same model (47.2 to 53.8) and a different corpus description. It cites eight source datasets and 46,469 guidelines, against five or seven datasets and 14,192 guidelines on the cards. The paper is older than the released weights (ev-paper).
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
ModerateApache-2.0 is declared and the weights are ungated and fine-tunable. The declaration sits in metadata only, with no LICENSE file, a paper that calls the corpus research-use, flagged source-dataset terms and a base acceptable-use policy of unclear reach, so the chain of title is unresolved.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
StrongWeights, a 533,888-row corpus with source fields and the generation, training and evaluation code are public on a fully open base (open_weights_recipe), with the unreleased intermediates and MOOVE vignettes listed in the README.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.