someone crammed a working AI language model onto a $10 chip. it writes actual stories

No cloud, no Wi-Fi, no GPU — just a cheap microcontroller quietly telling itself bedtime stories, one token at a time. Aliteq Lena Fischer · AI & Local Compute Editor

The ESP32-S3 has just 512KB of SRAM and 8MB of PSRAM — nowhere near what a language model normally needs.

The setup, in numbers

The model has 28.9 million parameters total. The previous record on ESP32-class hardware, using Andrej Karpathy's llama2.c architecture, topped out around 260,000.

The setup, in numbers

The trick is Google's Per-Layer Embeddings technique, first built for the Gemma 3n and Gemma 4 models: 25 million of those 28.9 million parameters live in a 16MB flash lookup table, not in RAM.

The setup, in numbers

It generates 9.5 tokens per second, trained entirely on the TinyStories dataset — so it writes toddler-length stories, not code or conversation.

The setup, in numbers

The full project is open-source on GitHub under an MIT license.

The setup, in numbers

It doesn't know anything, can't answer a question, and can't hold a conversation. It just writes toddler-length stories, forever, on a chip built for doorbells — and that's exactly the point.

Read the full story

Aliteq someone crammed a working AI language model onto a $10 chip. it writes actual stories Read the full story on Aliteq