esp-ai TinyStories

esp-ai TinyStories

A 29-million-parameter storyteller that runs entirely on your Attaky — no internet needed.

esp-ai TinyStories shows how much language model fits on one small chip. Switch it on and it writes a short English story on the screen, word by word, all on the device. Most of the model is read straight from flash memory instead of RAM, which is what lets a model this size run on the Attaky Core.

Highlights

• Fully offline: The whole model runs on the Attaky Core, with no Wi-Fi, account or cloud.
• Story on start-up: A short English story appears on the screen as the model writes it, about 10 words per second.
• Big model, small chip: 28.9 million parameters, with most of them kept in flash and read as needed.
• Speed report: When the story ends, the screen shows the model size and the measured speed.

Good to know

TinyStories has no input. Every start begins with "Once upon a time", and because the model always picks its most likely next word it tells the same story each time. It writes simple stories for young readers; it does not answer questions or know facts. For a model you can ask questions, see esp-ai Barista.

Compatibility

esp-ai TinyStories runs on Attaky Core; no other module is needed.

About this firmware

esp-ai TinyStories for Attaky is a port by Attaky of the TinyStories model and runtime from esp32-ai by Viacheslav Sierbov, adapted to the Attaky Core screen. The code and the model are released under the MIT License; the model was trained on the TinyStories dataset (CDLA-Sharing-1.0). Thanks to the upstream author.

Install

Install straight from the Attaky firmware store. Open this listing in your browser and flash it to your Attaky in one tap. There is nothing to download or install — the firmware store writes it to your device for you.