ESP32 AI projects

The ESP32 is the most common board on HardwareLab: 78 builds list it in their parts. Most use it as the device end of an AI system. The board captures audio, images or button presses, connects over Wi-Fi, and hands the heavy work to a model running in the cloud or on a computer at home. Voice assistants are the biggest group (21 builds), led by XiaoZhi, an open-source voice bot that talks to cloud language models and speech services, and Willow, whose README says: "Willow users can now self-host the Willow Inference Server for lightning-fast language inference tasks with Willow and other applications (even WebRTC) including STT, TTS, LLM, and more!"

A smaller group runs models on the chip itself, and those models are tiny by design. esp32-28m-llm generates text with a 28.9M-parameter model, needle-2-esp32 runs a 45M-parameter model offline on an ESP32-S3, and pfor-esp32-p4 fits a 180.9M-parameter mixture-of-experts model on an ESP32-P4. Narrow jobs work well on the board: esp-vision-detection does real-time object detection with no cloud, and esp32-ha-satellite-15 runs its wake word on the device and sends the rest to a home server.

So pick an ESP32 for small, cheap, always-connected devices: voice satellites, e-ink dashboards (6 builds), smart glasses (6) and desk robots (9). If you want a chatbot that runs fully offline, the builds here that do it use a Raspberry Pi or bigger hardware. Single-chip builds go further than you might expect: mimiclaw runs a personal agent on one ESP32-S3, and esp32-iphone-agent uses an ESP32-S3 board to let an AI model use a real iPhone.