tinyml-tensorlab - TI's MCU AI Toolchain

The Tiny ML Tensorlab repository is the starting point to install and explore TI's AI offering for MCUs. It supports Time Series Classification, Regression, Forecasting, Anomaly Detection, and Image Classification tasks across 20+ TI microcontrollers.

Detailed User Guide: TI Tiny ML Tensorlab User Guide

---

The TI Tiny ML ModelZoo: Contains the devices supported, a variety of models and all the applications worth exploring

---

Release History

• [2026-Feb] Release version 1.3.0 of the software

  Details

  • Device Support: 22 MCU devices supported:
    • AM1x: AM13E2
    • C2000 F28: F280013, F280015, F28003, F28004, F2837, F28P55, F28P65
    • C2000 F29: F29H85, F29P58, F29P32
    • MSP M0: MSPM0G3507, MSPM0G3519, MSPM0G5187
    • MSP M33: MSPM33C32,
    • Connectivity: CC2755, CC1352, CC1354, CC35X1,
    • AM26x: AM263, AM263P, AM261
  • Flows:
    • Timeseries Anomaly Detection flow supported
    • On Device Learning Mode Enabled
  • Applications Supported
    • 22 (4 generic + 18 specific applications)
  • Models:
    • 50+ generic models added over classification, regression, forecasting and anomaly detection tasks.
  • Model Optimization:
    • Partial Quantization Supported to enable best of precision and latency for regression models.
  • Compilation:
    • Upgraded TI MCU Neural Network Compiler for MCUs to 2.1.1 LTS

• [2025-Nov] Release version 1.2.0 of the software

  Details

  • Device Support:
    • Added MSPM0 based MCUs: MSPM0G3507, MSPM0G5187
    • Added Connectivity device: CC2745R10-Q1, CC2755R10
  • General:
    • Supports simple gain augmentation for classification tasks
    • Prints dataset file level confusion matrix for classification tasks
    • Golden Test Vectors for Regression tasks
    • Run modelmaker with only the config, no more target device required in the input.
  • Flows:
    • Timeseries Forecasting flows supported
    • L1, L2 normalization can be enabled in regression using lambda_reg param
  • Model Optimization:
    • How to use: Documentation updated.
    • Example code for performing regression in modeloptimization
    • Fixing clipping of input data to int8 or uint8 based on dataset (zero_point) (only the input zero point is fixed and not the intermediate layers)
    • BatchNorm is supported by GENERIC quantization
    • Experimental features like additional QDQ at input of model and floating bias can be enabled individually
    • Residual Add supported for different scales, zero points, but not optimised for TINPU
  • Compilation:
    • Upgraded TI MCU Neural Network Compiler for MCUs to 2.1.0 LTS
    • Supported all layer configs with 8-bit activations and 8-/4-/2-bit weights that can be offloaded to TI-NPU
    • Supported all layer configs with 8-bit activations and 8-bit weights that can be accelerated using the M33 Custom Datapath Extension (CDE).

• [2025-Aug] Release version 1.1.0 of the software

  Details

  • General:
    • Generic Timeseries Classification is available with fixed point reference dataset.
    • Compatible with C2000Ware 6.0.0
  • Model Optimization:
    • Aggressive Quantization Modes for Weights & Activation: 2W8A, 4W4A, 4W8A --> massive speedup and memory saved
    • Neural network Architecture Search for generating a TINPU compatible model directly based on user's dataset
  • Dataset:
    • Dataset can be split into train-test-val on a file-by-file basis or within-a-file basis
  • Device Support:
    • Full Support for F280013x
    • Preliminary Support for F29H85x and MSPM0G3507x
  • Compilation:
    • Upgraded TI MCU Neural Network Compiler for MCUs to 2.0.0
  • Windows Platform Specific:
    • Major quantization accuracy improvements
  • Miscellaneous:
    • Fixed model performance data that appears on the terminal when a training is initiated
    • Added Model Descriptions for all models
    • Setup of the repos is now smoother and cleaner

• [2025-Apr] Major feature updates (version 1.0.0) of the software

  Details

  • General:
    • Tiny ML Modelmaker is now a pip installable package!
    • Existing models can be modified on the fly through a config file (check Tiny ML Modelmaker docs)
    • MPS (Metal Performance Shaders) backend support for Mac host devices!
  • Technology:
    • PTQ and QAT flows supported in tinyml-modelmaker, tinyml-modeloptimization
    • Ternary, 4 bit Quantization support in tinyml-modelmaker
  • Flows:
    • Regression ML tasks supported
    • Autoencoder based Anomaly Detection task supported
  • Feature Extraction:
    • Feature Extraction transforms are now modular and compatible with C2000Ware 5.05 only
    • Supports Haar and Hadamard Transform
    • Golden test vectors file has one set uncommented by default to work OOB
  • Data Visualisation:
    • Multiclass ROC-AUC graphs are autogenerated for better explainability of reports and help select thresholds based on false alarm/ sensitivity preference
    • PCA graphs are auto plotted for feature extracted data - Helps in identifying if the feature extraction actually helped
    • Run now begins with displaying inference time, sram usage and flash usage for all the devices for any model.
  • Dataset
    • Goodness of Fit of dataset now enabled.
  • Extensive Documentation & Know-How Examples to use Modelmaker

• [2024-November] Updated (version 0.9.0) of the software

• [2024-August] Release version 0.8.0 of the software

• [2024-July] Release version 0.7.0 of the software

• [2024-June] Release version 0.6.0 of the software

• [2024-May] First public release (version 0.5.0) of the software