Overview of image classification STM32 model zoo

January 22, 2026 · View on GitHub

The STM32 model zoo includes several models for image classification use cases pre-trained on custom and public datasets. Under each model directory, you can find the following model categories:

  • Public_pretrainedmodel_public_dataset contains public image classification models trained on public datasets.
  • ST_pretrainedmodel_custom_dataset contains different image classification models trained on ST custom datasets using our training scripts.
  • ST_pretrainedmodel_public_dataset contains different image classification models trained on various public datasets following the training section in STM32 model zoo.

Feel free to explore the model zoo and get pre-trained models here.

Model Families

You can get comprehensive footprints and performance information for each model family following the links below:

Mobile-Optimized Architectures

You can get footprints and performance information for each model following links below:

  • mobilenetv1 – Efficient depthwise separable convolutions (alpha variants: 0.25, 0.50, 1.0)
  • mobilenetv2 – Inverted residual blocks with linear bottlenecks (alpha variants: 0.35, 1.0, 1.4)
  • fdmobilenet– Fast downsampling variants for reduced latency (two variants: 0.25 and in-house designed model)

STMicroelectronics In-house Model

  • efficientnet - A ST customization of first published version of EfficientNet tailored for STM32 platforms, and modified to be quantization-friendly
  • st_mnistv1 - A ST customized topology leveraging the benefits of depthwise separable convolutions and well suited for MNIST-like datasets.

Standard Architectures

  • efficientnetv2 - One of the best topology for image classification (several variants: B0 (224x224), B1 (240x240), B2 (260x260), B3 (300x300), S (384x384))
  • squeezenetv11 – Fire modules with squeeze and expand
  • resnet - ResNetv1-8 model trained on CIFAR-10 and CIFAR-100 datasets
  • resnet50v2 - Well known residual connection architecture to target more challenging image datasets

To get started, update the user_config.yaml file, which specifies the parameters and configuration options for the services you want to use. The model section of this yaml specifically relates to the model definition. Some topologies are already registered and can be accessed by the model_name attribute. The exhaustive list of possible values is provided hereafter:

  • 'custom_model'
  • 'st_efficientnetlcv1'
  • 'st_fdmobilenetv1'
  • 'st_mnistv1'
  • 'efficientnetv2b0'
  • 'efficientnetv2b1'
  • 'efficientnetv2b2'
  • 'efficientnetv2b3'
  • 'efficientnetv2s'
  • 'fdmobilenet_a025'
  • 'fdmobilenet_a050'
  • 'fdmobilenet_a075'
  • 'fdmobilenet_a100'
  • 'mobilenetv1_a025'
  • 'mobilenetv1_a050'
  • 'mobilenetv1_a075'
  • 'mobilenetv1_a100'
  • 'mobilenetv2_a035'
  • 'mobilenetv2_a050'
  • 'mobilenetv2_a075'
  • 'mobilenetv2_a100'
  • 'mobilenetv2_a130'
  • 'mobilenetv2_a140'
  • 'resnet50v2'
  • 'resnet8'
  • 'resnet20'
  • 'resnet32'
  • 'squeezenetv11'

Quick Selection Guide

By Inference Speed (on STM32N6570-DK, dataset food101, input size: 224x224)

  • Ultra-Fast (<5ms): fdmobilenet_a025 (1.29ms), st_fdmobilenetv1 (1.67ms), mobilenetv1_a025 (2.37ms)
  • Fast (5-10ms): mobilenetv1_a050 (5.38ms), mobilenetv2_a035 (5.43ms), squeezenetv11 (7.97ms)
  • Moderate (10-20ms): mobilenetv1_a100 (16.36ms), mobilenetv2_a100 (16.43ms), st_efficientnetlcv1 (17.31ms)
  • Balanced (20-40ms):
  • Large (>40ms): efficientnetv2b0 (57.05ms), efficientnetv2b1 (80.50ms), efficientnetv2b2 (140.38ms), resnet50v2 (238.49ms)

By Model Size (weights)

  • Tiny (<500KB): fdmobilenet_a025 (148KB), st_fdmobilenetv1 (167KB), mobilenetv1_a025 (241KB), mobilenetv2_a035 (423KB)
  • Small (500KB-1.5MB): squeezenetv11 (753KB), mobilenetv1_a050 (865KB)
  • Medium (1.5-3MB): mobilenetv2_a100 (2336KB)
  • Large (>3MB): mobilenetv1_a100 (3348KB), efficientnetv2b0/b1/b2 (4237KB to 6885KB), resnet50v2 (13268KB), efficientnetv2S (14837KB)

By RAM Requirements (on STM32N6570-DK, dataset food101, input size: 224x224)

  • Internal RAM Only (<1MB): Most MobileNet v1 and v2, st_efficientnetlcv1, st_fdmobilenetv1...
  • Requires External RAM: efficientnetv2 b2/b3/s (528KB to ~3500KB), resnet50v2 (2352KB)

Platform Support

  • Primary Target: STM32N6570-DK with NPU acceleration. The smallest can also be considered for STM32H7.
  • Framework: some models are avaliable in tflite full Int8, others are available in ONNX QDQ full Int8 or mixed-precision (Int8/Int4). Activations are always in Int8.
  • STEdgeAI Core: Version 3.0.0+

Performance Notes

  • All inference times measured on STM32N6570-DK with NPU/MCU execution
  • Models using only internal RAM are preferred for simpler deployment
  • External RAM models (some efficientnetv2 and resnet50v2) offer different accuracy/performance tradeoffs

Feel free to explore the model zoo and get pre-trained models here.

For training and deployment guidance, refer to the STM32 AI model zoo documentation.