A 28.9M‑parameter language model now runs on an ESP32‑S3 microcontroller that costs roughly $8. The firmware executes entirely on the chip, with no server connection, and streams each generated word to a tiny screen at about 9 tokens per second. The last chip‑based model held only 260,000 parameters, so this marks a ~100‑fold increase.

The breakthrough comes from storing most of the 25 million‑parameter embedding table in the device’s 16 MB flash, while the 512 KB SRAM hosts only the small, compute‑heavy portion of the network. Google’s Per‑Layer Embeddings design from the Gemma family enables the model to fetch only ~450 B per token, keeping latency low and memory usage minimal.

Trained on the Tiny Stories dataset created by Ronen Eldan and Yuanzhi Li at Microsoft Research, the model excels at writing short, coherent tales but cannot answer questions or write code, reflecting the limited reasoning capacity of the small compute core. The firmware, training scripts, and full results are publicly released under MIT license.

This work demonstrates that a large LLM can be embedded on a budget microcontroller, highlighting the potential for edge‑AI applications in constrained environments.