Developer runs 28.9M-parameter model on $10 microcontroller

A developer got a 28.9-million-parameter language model running on a $10 ESP32-S3 microcontroller. The setup uses Google's Per-Layer Embeddings technique and stores its embedding table on 16MB of flash memory to fit within the chip's limited RAM.

Why it matters: It's a concrete demonstration of how far model compression can push language models onto extremely cheap edge hardware, well below the microcontrollers typically discussed for on-device AI. That has implications for cost-sensitive embedded applications like sensors, toys, and appliances that previously couldn't justify any onboard language capability.