Qapla': Training AI on ESP32 microcontroller Qapla' is an open-source… — Embedded Systems

Embedded Systems@embedded_system · 14.1K subscribers Open in TelegramOpen

Qapla': Training AI on ESP32 microcontroller

Qapla' is an open-source project that demonstrates training a transformer model entirely from scratch on an inexpensive ESP32-S3 microcontroller.

Instead of running a pre-trained model, it performs the complete training process—including backpropagation and weight updates—directly on the device.

Features of the project:

Full on-device transformer training

Handwritten C backpropagation

No pre-trained model required

Open and reproducible setup

Qapla' proves that even low-cost microcontrollers can train small transformer models without GPUs or cloud infrastructure. It is an interesting project for embedded AI developers, TinyML enthusiasts, and researchers exploring on-device learning under extreme hardware constraints.