Five ESP32 Builds Push Local AI From a Dev Board to a Drone's Camera

A 28.9M-Parameter LLM Lands on an ESP32-S3

A hobbyist going by ninhnn2 published a from-scratch commit on October 21, 2026 for PLE TinyLM, a 28.9-million-parameter language model built to run on an ESP32-S3. The trick is keeping most parameters as per-layer embeddings in flash instead of RAM, so the same compact model.bin file runs on the ESP32-S3 in Xtensa C, on a desktop CPU in portable C, and on a Jetson over CUDA, according to the commit log.

This is a first release, not a benchmark drop: no tokens-per-second figures are published yet, and the project leans on a self-authored Vietnamese-language course on roofline analysis and quantization rather than a peer-reviewed paper. Still, one binary format spanning a microcontroller and a GPU board is useful plumbing for anyone building a model once and deploying it from an $8 chip up.

A $7 ESP32 Gives an LLM Hands on a Real iPhone

esp32-iphone-agent, published by developer onehumanbeing on October 11, 2026 and past 3,400 stars by October 21, turns a roughly $7 ESP32 board into hands and eyes for an iPhone: AirPlay streams the screen in, USB HID sends taps and gestures back out, and any OpenAI-compatible endpoint supplies the decisions, according to the project repo.

The honest caveat: the ESP32 here is a hardware bridge, not where the model runs. Point the endpoint at a cloud API and this is a cloud brain with a cheap body; point it at a local llama.cpp or Ollama server on the same network and the control loop stays on local hardware. Either way, it is a cheap way to let a model operate a real phone instead of a simulator.

A Knob-or-Wake-Word Voice Assistant, Model Left Open

raullenchai's esp32-thinks, created October 12, 2026 and up to 236 stars by October 19, is a personal voice assistant built around an ESP32 that listens for a wake word or a physical knob press, captures audio locally, and routes it to any OpenAI-compatible model backend, per the repo.

As with the iPhone-agent project, how much of this stays on-device depends on where the user points the endpoint. Wake-word detection and audio capture happen on the chip; the reasoning can be a small local model or someone else's cloud API. It is a solid template for builders who want the hardware fixed now and the model choice left for later.

ESP-Mosaico Pairs a New ESP32-S31 With a Touchscreen for AI-Assisted Builds

CNX Software reported on October 21, 2026 on ESP-Mosaico, a modular development platform built around the new ESP32-S31 with a 480 x 480-pixel OLED touchscreen, haptic feedback, a built-in battery, and USB Type-C, per the writeup.

Billed for AI-assisted hardware creation, the board targets a workflow rather than an inference job: letting a coding agent iterate on firmware and UI for a battery-powered touchscreen gadget without a builder touching a soldering iron for every change. No model runs on it out of the box; the AI here is in the tooling, not the silicon.

An ESP32's Hidden Radio Starts Flying 5 GHz Drone Video

Hackaday covered a new application on October 20, 2026 for the ESP32's recently revealed software-defined radio facility: streaming live video over 5 GHz from a drone, built on the transmit capability unlocked only a week earlier, per the article.

The caveat is range and robustness: this is a stock Wi-Fi radio pressed into service as a video link, not a dedicated analog or digital FPV system, and the write-up treats it as an early proof of concept rather than a flight-ready product. Still, it shows how much latent radio hardware sits unused inside an ordinary ESP32 module.

Five builds, one chip family: the ESP32 keeps turning up as the cheapest way to give an AI agent a body, a microphone, or a radio, even when the model itself lives somewhere else entirely.