README.md

metadata

title: README
emoji: ⚡
colorFrom: red
colorTo: gray
sdk: static
pinned: false

Innovating with edge AI on STM32 and Hugging Face.

STMicroelectronics is a global semiconductor leader pushing artificial intelligence down to the most resource-constrained microcontrollers. With the STM32 AI ecosystem, ST provides an end-to-end pipeline — from pre-trained models in the Model Zoo to bare-metal optimized deployment — enabling embedded developers to build intelligent applications without deep ML expertise. Models are optimized, quantized and validated to run directly on ST Neural-ART but also Cortex-M4, M7, M85 and M33 cores.

---

End-to-End AI Pipeline

+----------------------------+
|          EXPLORE           |
+----------------------------+
|     STM32 AI Model Zoo     |
+----------------------------+
              |
              v
+----------------------------+
|           TRAIN            |
+----------------------------+
|  STM32 AI Model Zoo        |
|  Services                  |
+----------------------------+
              |
              v
+----------------------------+
|     OPTIMIZE / QUANTIZE    |
+----------------------------+
|  STM32 AI Model Zoo        |
|  Services                  |
+----------------------------+
              |
              v
+----------------------------+
|     EVALUATE / PREDICT     |
+----------------------------+
|  STM32 AI Model Zoo        |
|  Services                  |
+----------------------------+
              |
              v
+----------------------------+
|         BENCHMARK          |
+----------------------------+
|  STM32Cube AI Studio       |
|  STM32 Developer Cloud     |
+----------------------------+
              |
              v
+----------------------------+
|          CONVERT           |
+----------------------------+
|  STM32Cube AI Studio       |
|  ST Edge AI Core           |
+----------------------------+
              |
              v
+----------------------------+
|           DEPLOY           |
+----------------------------+
|  STM32Cube ecosystem       |
|  (tools, middleware, BSP)  |
+----------------------------+

This diagram summarizes the typical STM32 edge AI workflow from model discovery to on-device deployment:

1. Explore: Start from the STM32 AI Model Zoo to browse available architectures, pretrained checkpoints, and application examples.
2. Train: Use Model Zoo Services to retrain an existing model or build a task-specific pipeline on your own dataset.
3. Optimize / Quantize: Reduce model size and compute cost so the network fits embedded constraints while preserving the best possible accuracy.
4. Evaluate / Predict: Validate accuracy, inspect predictions, and compare tradeoffs before moving to hardware execution.
5. Benchmark: Measure latency, memory footprint, and target compatibility with STM32Cube AI Studio and STM32 Developer Cloud.
6. Convert: Transform the trained model into STM32-ready artifacts using STM32Cube AI Studio and ST Edge AI Core.
7. Deploy: Integrate the generated code into the STM32Cube ecosystem, including firmware, middleware, and board support components.

In short, the flow shows how a model moves from selection and training to optimization, hardware validation, and final integration on STM32 devices.

Build, Optimize and Deploy AI/ML on STM32

• STM32 AI Model Zoo: A GitHub collection of reference machine learning models optimized for STM32 microcontrollers.
• Application-Oriented Model Library: A large set of models ready for re-training across multiple use cases.
• Pre-trained Models Across Frameworks: Reference models variants available for PyTorch, TensorFlow, and ONNX workflows.
• End-to-End Scripts & Services: Tools to retrain, quantize, evaluate, and benchmark models on custom datasets, plus autogenerated application code examples via stm32ai-modelzoo-services
• Fast Deployment + Full Customization: Use pretrained categories for quick deployment, or apply transfer learning / full training from scratch on your own data.
• Reference Performance Metrics: Results provided on STM32 MCU, NPU, and MPU targets for both float and quantized models.
• Expanded Framework Support: Comprehensive PyTorch support complements TensorFlow and ONNX in unified end-to-end workflows (train, evaluate, quantize, benchmark, deploy).

---

Key Tools & Ecosystem

• STEdgeAI Core: Converts trained neural networks into optimized C code for STM32.
• STM32 AI Model Zoo services: This repository provide scripts and workflows to ease end-to-end AI model training and integration on ST devices. They offer a valuable foundation to add AI capabilities to STM32-based projects.
• STM32 AI Model Zoo The repository with a of reference pre-trained machine learning models optimized for STM32 microcontrollers generated thanks to the STM32 AI Model Zoo services.
• Integration with Popular Frameworks:
  • TensorFlow / Keras
  • PyTorch (via ONNX export)
  • ONNX Runtime pipelines

---

Links

• STM32 AI Model Zoo services
• STEdgeAI Core
• STM32 Developer Cloud
• STM32AI Model Zoo
• STM32AI Cube Studio

---

🤝 Contact & Contributions

• For technical questions: ST EdgeAI Community
• For issues or feature requests, use the Issues or Discussions tabs in the respective repos.
• Contributions and feedback on models, pipelines, and docs are welcome.