esp32-ai
Runs a 28.9M-parameter language model entirely on an ESP32 microcontroller, no network needed.
This project runs a 28.9-million-parameter language model directly on an ESP32-S3 microcontroller, generating text at 9.88 tokens/second on a small attached screen with nothing sent to a server. It fits by keeping 25 million of those parameters in flash rather than RAM — a Google Gemma 3n technique called Per-Layer Embeddings — pulling only about 450 bytes per token from a flash lookup table instead of loading the whole thing into memory. The model is trained on TinyStories, so it writes short, coherent stories; the README states plainly it will not answer questions, follow instructions, write code, or know facts.
What holds up
- +Runs fully offline on the chip — nothing is sent to a server, by hardware design, not just a setting.
- +Model download is verified against a pinned SHA-256 and size before install, rejecting a tampered download.
- +Two ready-made models (TinyStories, Barista for espresso Q&A) share the same two-command fetch/deploy flow.
Mind the limits
- −Will not answer questions, follow instructions, write code, or know facts — only generates short stories.
- −Requires actual ESP32-S3 hardware — not something you can try in a browser or on a different chip.
Featured here? Take the badge
Put it on your site — it links back to this review. Free for every listed product, always.
<a href="https://stillworks.dev/products/p/esp32-ai/"><img src="https://stillworks.dev/badge/esp32-ai.svg" alt="Picked by StillWorks" width="250" height="54"></a>