Edge AI Development Tools

Nordic ships a layered tool stack for Edge AI: a high-level no-code lab for common tasks, an SDK add-on for production integration, a compiler that produces NPU-ready model headers, and a partnership with Edge Impulse for end-to-end MLOps. You can mix and match — most production teams use the Edge AI Lab for prototyping and the Edge AI Add-on + Axon Compiler for shipping firmware.

Edge AI Add-on for nRF Connect SDK (v2.0)

The Edge AI Add-on is the official Nordic extension to the nRF Connect SDK that brings ML support into your Zephyr application.

v2.0 (2026) added:

• Low-level Axon NPU drivers
• Axon support inside the nRF Edge AI library (nrf_edgeai_lib)
• The Axon NPU Compiler for converting LiteRT models
• Compatible code samples (KWS, audio classification, sensor analytics)

v1.0 (December 2025) previously added support for custom Neuton models inside nRF Connect SDK.

You install the add-on into your existing nRF Connect SDK workspace; it appears as additional Kconfig options and Zephyr modules under CONFIG_NRF_EDGEAI_LIB=y.

Axon Compiler

The Axon Compiler converts a trained model into a header file the firmware can include directly.

Because the output is a plain C header, integration into a Zephyr build is a one-line #include plus a call into nrf_edgeai_lib. There's no runtime model loader and no dynamic memory allocation — everything lives in NVM and is mapped into the inference engine at boot.

Nordic Edge AI Lab

A cloud-based, no-code environment for building Axon-ready models without setting up a Python ML toolchain locally.

Text-to-wake-word

Type the wake phrase you want, and the Lab synthesises a training corpus, trains a small KWS model, and produces an Axon-compiled .h you can drop straight into your firmware. This is the fastest path from "I want my device to listen for 'Hey Polaris'" to a flashable build.

Model builder for the Axon NPU

For other workloads (audio classification, sensor anomaly detection), the Lab provides a guided model-building flow that picks an architecture that maps cleanly to the Axon's natively accelerated op set, then runs the same Axon Compiler under the hood.

Edge Impulse integration

If your team already uses Edge Impulse for data collection, labelling, and MLOps, the nRF54LM20 DK is supported as a first-class target. You can:

• Stream sensor data from the DK directly into an Edge Impulse project
• Train classifiers / regressors / anomaly detectors in the Edge Impulse Studio
• Deploy the resulting model as a C++ library or, more importantly for Axon, as a quantised .tflite you then run through the Axon Compiler

See docs.edgeimpulse.com/hardware/boards/nordic-semi-nrf54LM20-dk.

Neuton custom models

For workloads where the NPU is overkill (low-rate IMU / PPG, simple sensor analytics), Nordic supports Neuton — a partner ML platform that generates extremely small, CPU-run models suited to the Cortex-M33 alone. Neuton models work on any nRF54L Series part (not just the LM20B) and were the first ML integration shipped via the Edge AI Add-on v1.0.

Pick Neuton when:

• Your input is low-rate (e.g. < 100 Hz accelerometer)
• Your model is small (a few KB)
• You don't need the NPU and want the model to run on parts without one

Pick the Axon NPU when:

• You're doing audio (KWS, classification) or higher-rate sensor analysis
• Inference latency or per-inference energy matters
• Your model uses convolutions, depthwise conv, or dense layers

How FirmwareMaestro composes these

When you generate a project for the nRF54LM20B + Edge AI use case, FirmwareMaestro produces:

• A prj.conf enabling CONFIG_NRF_EDGEAI_LIB, the Axon driver, and any DSP preprocessing modules your model needs (PDM mic, FFT, MFCC)
• A Devicetree overlay for the audio / sensor source on the LM20 DK
• A main.c that initialises the NPU, loads the compiled model header, and runs inference inside a dedicated Zephyr work queue with BLE reporting
• An Architecture document that records the model footprint, inference budget, and pre-processing chain
• A HAL document describing the audio / sensor capture path
• A State Machine for "idle → trigger → infer → report → sleep"

The compiled model header itself comes from the Axon Compiler or the Edge AI Lab — FirmwareMaestro doesn't train models, it integrates them.

Where to next

• See the silicon details: nRF54LM20B SoC
• Understand the accelerator: Axon NPU architecture
• Walk through a build: Getting started