ModusToolbox™ Machine Learning Middleware Library Release Notes

July 21, 2026 · View on GitHub

The ModusToolbox™ ML Middleware Library is a set of helper functions to handle the Machine Learning models and capability to receive stream validation data.

Please refer to the README.md for a complete description of the ModusToolbox™ ML Middleware Library.

What's Included?

This release of the ModusToolbox™ ML Middleware library includes support for the following functionality:

utils - functions to help handling the inputs/outputs of a machine learning model
model - functions to initialize and inference a ML model
stream - functions to stream validation data

Support floating-point and fixed-point variants:

32-bit floating-point
16-bit fixed-point input
8-bit fixed-point input
8-bit fixed-point weight

The MTB-ML Middleware library supports Windows, Linux and Mac OS.

What Changed?

v3.4.0

Added macOS arm64 native support for the middleware streaming application (no Rosetta required)

v3.3.0

Fixed FullyConnected (FC) per-channel quantization in NNLite integration
Fixed dequantization pointer arithmetic in mtb_ml_utils_model_dequantize for correct handling of varying output data type sizes

v3.2.0

Implemented Multi-Input/Multi-Output (MIMO) model support. It is available in both TFLM (with interpreter) and TFLM_LESS (pre-interpreted) configurations
Runtime-quantization params usage added - quantization can be defined both design-time and run-time
Added RELU_0TO1 and RELU_N1TO1 activation function support
Functional fixes for U55 cycle count and cache management

v3.1.0

Implemented support for PSOC Edge™ capabilities:
- Cortex-M33 (CPU-based inference)
- NNLite™ NPU integration (Cortex-M33 core only)
- EthosU U55 NPU integration (Cortex-M55 core only)
Added LSTM models support
Implemented NNLite™ NPU RToS integration
Implemented EthosU U55 NPU RToS integration
Implemented EthosU U55 NPU cycle count profiling
MTB-HAL Support added
Support of ml-inference obsoleted

v2.0.1

Added mtb_ml_utils_model_quantize & mtb_ml_utils_convert_tflm_flt_to_int8 functions

v2.0.0

Support for TensorFlow Lite Micro
PSOC6™ as a target platform (Cortex-M4)

v1.0.0

Initial release

Supported Software and Tools

This version of the was validated for compatibility with the following Software and Tools:

Software and Tools	Version
ML TFLiteU	3.4.0
ML Coretools	3.3.0
GCC Compiler	14.2
ARM Compiler 6	6.22
LLVM Toolchain	19.1.5

Minimum required ModusToolbox™ Software Environment: v3.6

More information

Use the following links for more information, as needed: