⚡ qapla - Train AI On A Microchip

A character-level transformer that trains itself on an ESP32-S3 — full backprop, written in C, running on $8 hardware.

🚀 Getting Started

Welcome! qapla is a unique project that lets you train a real transformer model directly on an ESP32-S3 microcontroller. No cloud, no GPU — just a tiny chip doing the heavy lifting. Follow the steps below to begin.

📥 Download & Install

Step 1: Visit the download page to get the application.

🔗 Go to Download Page

Step 2: Once on the page, find the latest release and download the Windows installer (it will be a file ending in .exe).

Step 3: Double-click the downloaded file and follow the setup wizard. qapla will install and create a shortcut on your desktop.

Step 4: Launch qapla from the shortcut or Start Menu. The interface will guide you through connecting your ESP32-S3 and starting a training session.

💡 What You Can Do

Train a small transformer model on your own text data (characters, words, or patterns).
Watch real-time loss curves and token predictions as the model learns on-device.
Export the trained weights and run inference on the ESP32-S3 without a computer.
Experiment with learning rates, sequence lengths, and other hyperparameters — all from a simple interface.

🔧 Requirements

Windows: 10 or 11 (64-bit).
Hardware: An ESP32-S3 development board with USB connection.
USB driver: CP210x or CH340 driver installed (usually automatic).
Storage: 100 MB free space for the software.

📖 How It Works

qpla implements a full training loop from scratch in C — forward pass, cross-entropy loss, backpropagation, and optimizer update — all running on the ESP32's dual-core processor. The host computer acts as a terminal and data feeder. You provide a text file, and the chip learns to predict the next character, byte by byte.

❓ Troubleshooting

Board not detected? Ensure the USB cable supports data (not just charging). Try a different port.
Training stalls? Reset the board by pressing the EN button, then restart qapla.
Errors on launch? Run as Administrator or install the latest Visual C++ Redistributable.

🔒 Privacy & Security

All training happens locally on your ESP32 and computer. No data leaves your machine. qapla does not phone home or collect any information.

📬 Support

Issues or questions? Visit the releases page for links to documentation and community forums.