Release model
This page describes the version identifiers, release artifacts, and compatibility rules for the Edge AI Add-on, nRF Connect SDK, Axon NPU software, and the nRF Edge AI Library.
Add-on release identifiers
The version string is defined in the VERSION file at the repository root.
Git tags use a leading v prefix, for example v3.0.0.
Use the tagged add-on version together with the nRF Connect SDK revision listed in west.yml.
Do not mix an add-on tag with a different nRF Connect SDK version unless a release note or changelog explicitly states support for that combination.
Component version streams
The following table summarizes the version identifiers:
Component |
Version source |
Where to find it |
|---|---|---|
Edge AI Add-on add-on |
|
Git tag |
nRF Connect SDK |
|
|
Axon NPU software |
Independent Axon release, for example |
Axon NPU changelog, driver and compiler binaries in the add-on tree |
nRF Edge AI Library |
Independent library release, for example |
|
Nordic Edge AI Lab solutions |
Solution export version embedded in generated model sources |
|
Axon NPU compatibility
Axon NPU software consists of the on-device driver, the host-side TFLite compiler, and the compiled model artifacts they produce.
Backward compatibility (older models, newer driver)
Compiled Axon models built with an older compiler and driver generation generally run on a newer Axon driver in the same major feature set. When a compiler or driver fix changes numerical behavior, recompile the affected models with the compiler version shipped in your target add-on release.
Since the Axon release v1.2.0, compiled model headers can declare a minimum supported Axon version. If a model requires driver features that are not present, the build or initialization fails instead of running with undefined behavior.
Forward compatibility (newer models, older driver)
Newer compiled models may run on an older Axon driver only when they do not depend on features introduced after that driver version. When in doubt, compile the models with the Axon compiler version bundled in your selected Edge AI Add-on tag and run them with the matching driver from the same tag.
nRF Edge AI Library compatibility
The nRF Edge AI Library integrates the Neuton CPU models and Axon NPU models through a shared runtime API. Each library release lists the Nordic Edge AI Lab solution versions and Axon driver versions it supports in the nRF Edge AI Library changelog.
Runtime compatibility between library and model
Models exported from Nordic Edge AI Lab embed a solution runtime version.
At startup, applications should verify that the on-device library matches that export using the nrf_edgeai_is_runtime_compatible() function.
If the check fails, regenerate or redeploy the model for the nRF Edge AI Library version in your add-on release instead of forcing a mismatch at runtime.
This check is also performed by the nrf_edgeai_init() function.
Backward compatibility (older models, newer library)
Within the supported ranges published in the nRF Edge AI Library changelog, a model generated for an older Nordic Edge AI Lab solution version is expected to work with a newer nRF Edge AI Library from a later add-on release. Patch and minor library releases focus on fixes and added capabilities, such as support for additional Axon driver versions.
When the release notes or the library changelog call out API or pipeline changes, regenerate the model sources from Nordic Edge AI Lab before upgrading the add-on in a product firmware line.
Forward compatibility (newer models, older library)
A model generated for a newer Nordic Edge AI Lab solution version is not guaranteed to run on an older nRF Edge AI Library.
Using the nrf_edgeai_is_runtime_compatible() check function is the supported way to detect this condition.
Axon-backed nRF Edge AI Library releases also document which Axon driver versions they support.