Context card
Model version, software version, device version: three numbers that are not the same
What is the difference between the model version, the software version and the device version of an AI-enabled medical device?
The short answer
The model version identifies the trained model (weights, training data snapshot, hyperparameters), the software version the released device software that contains it, and the device version the product as placed on the market, identified by its UDI. One software release can carry a new model; a new model can require a new UDI-DI. After a model update all three must stay traceable to each other and to their validation evidence.
For: Software, regulatory and documentation teams of AI-enabled medical devices
Key points
- The model version names a trained artefact; it is set in the data science tools.
- The software version names the released software item under IEC 62304 configuration management.
- The device version is what the UDI identifies and what the file and labelling describe.
- Under the MDR, a change to the software that modifies the original performance or intended use requires a new UDI-DI.
- The file needs a mapping: which model in which software in which device, validated by which evidence.
The context
Three versions, three owners
- Model version — the trained model: the weights, the snapshot of training, tuning and test data, the hyperparameters and the code that produced them. It lives in model registries and data science tools. A model card describes it.
- Software version — the released software item that contains the model, under the configuration management of IEC 62304. It is what verification and validation are run against.
- Device version — the product as placed on the market, identified by its UDI (UDI-DI for the model or version, UDI-PI for the production unit) and registered in EUDAMED.
Why they drift apart
A data science team can produce ten model versions between two software releases. A software release can change the user interface without touching the model. The MDR rules on UDI assignment (Annex VI Part C) require a new UDI-DI when a software change modifies the original performance, safety or intended use; minor revisions get a new UDI-PI. That is a regulatory decision, not a version number in a repository.
What must stay traceable after a model update
- Which model version is in which software version, and which software version is in which device version.
- The data the model was trained, tuned and tested on, with its provenance (see training data provenance).
- The verification and validation of the whole device with that model.
- The risk assessment of the change and, in the USA, whether it falls under an authorised PCCP.
- The labelling and IFU text that describes the performance of that version.
What this means for the technical writer
An IFU that states a sensitivity or specificity states it for a version. If the model changes and the number changes, the IFU of that device version changes — and older versions keep their text. Documentation that cannot name the version it describes cannot be used to answer a question about a specific device.
Questions readers ask next
- Does every model update need a new UDI-DI?
- Not every one. Under the MDR, a new UDI-DI is required when a software change modifies the original performance, safety or intended use of the software or the interpretation of data; minor revisions such as bug fixes get a new UDI-PI. Check the rules of MDR Annex VI Part C and the MDCG guidance for the case.
- Is the model a SOUP item?
- A model developed by the manufacturer is not. A third-party model used inside the device can be software of unknown provenance under IEC 62304 and must be handled as such.
Sources
- Regulation (EU) 2017/745 on medical devices (MDR) — EUR-Lex, 5 May 2017Law in force: Binding legislation that applies as of the review date.
- IEC 62304:2006+A1:2015 Medical device software — Software life cycle processes — IEC / ISO, June 2015Standard: A published standard or specification. Not cited in the Official Journal as a harmonized standard, so it gives no presumption of conformity on its own.
- Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (final guidance, revised August 18, 2025) — U.S. Food and Drug Administration, 4 December 2024Final guidance: Guidance published in its final form by an authority or an official body. Guidance explains the law; it is not law itself.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) — EUR-Lex, 12 July 2024Law in force: Binding legislation that applies as of the review date.
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.