Needle 2 Esp32
45M-parameter LLM offline on an ESP32-S3 turning English into valid commands.
- Est. cost
- Cost unknown
- Posted
- GitHub stars
- 193 (GitHub)
Overview
A 45-million-parameter language model doing grammar-guaranteed tool calling entirely on one ESP32-S3: type a request and the board's RGB LED does what you asked, with no network.
How it works
An independent C99 engine (about 3,700 lines) reads the model weights from 16 MB of flash and uses 8 MB of octal PSRAM for its KV cache. Output is constrained so every tool call is valid JSON. It runs at about 1.87 tokens per second, roughly 25 seconds per request with reasoning turned off.
Hardware
- ESP32-S3
This page is a short summary written by HardwareLab. Build steps, code and photos belong to the creator; follow “View original” for the full project.
More in Local LLM Rigs & Edge Agents
See all 17Esp32 Iphone Agent
A $7 ESP32 lets an AI drive a real iPhone via AirPlay and USB HID.
- ESP32 (USB HID)
- iPhone
IntermediateCost unknown1.4KPocket Tank
A 14M-parameter LLM keeps a virtual fish tank alive on an $8 ESP32.
- ESP32-S3
AdvancedCost unknown207ESP32s3 LLM Cluster
Seven ESP32-S3s daisy-chained to run a 0.4B ternary LLM.
- 7x ESP32-S3 over SPI
AdvancedCost unknown234