A language model on an ESP32-S3 An $8 chip, no internet. The board continues your story token by token, in real time. Source on GitHub, full report.

1. Flash the board (once)

Needs an ESP32-S3 with 16 MB flash and 8 MB octal PSRAM (for example ESP32-S3-DevKitC-1 N16R8), a USB cable and Chrome or Edge on a computer. Flashing erases the board.

Your browser cannot flash boards. Use Chrome or Edge on a desktop. Flashing needs HTTPS.

Designed for bytes per token: 0.63 bits/byte on TinyStories, ~80-95 tokens/s.

The slvDev/esp32-ai TinyStories model (MIT) on our engine: ~87 tokens/s vs 10.6 on its own firmware.

After flashing, close the flasher window and reload this page (F5) so it releases the port, then press Connect board below. If the board is not in the port list, use the USB connector marked COM or UART.

2. Write with it

Board not connected
Style

The model was trained on children's stories (TinyStories), so it continues stories; it does not answer questions. Everything runs on the board: the page only turns text into token ids and back.

—tokens per second
—ms per token
0tokens written
—ms to read the prompt
Board output
Connect the board and press Write.
Time per token, ms