TinyONNX · Mini Edge Runtime

A tiny, dependency-free machine learning runtime built from scratch for learning and edge inference.

Overview · Hardware · Quick start · Architecture · Supported operations · Development

---

Overview

TinyONNX is a compact implementation of an ONNX-inspired model format and inference runtime for mobile, edge, and on-device AI. The project keeps the entire execution path visible: tensors, graph validation, operator schemas, planning, kernels, and native execution.

It does not depend on ONNX, ONNX Runtime, NumPy, or another machine learning framework.

Important

TinyONNX is an educational, experimental runtime. Its JSON model format is inspired by ONNX architecture, but it is not the official ONNX wire format and is not intended for production workloads yet.

Hardware support

Board	Image	Progress	Documentation / example
ESP32-CAM		Planned; not yet validated.	Not available yet.
ESP32 DevKit		Complete reference: ESP-IDF build and flash verified.	Build guide · Firmware example
ESP32-S3		Planned; not yet validated.	Not available yet.

License

Distributed under the MIT License.

_{Built to make edge inference internals small enough to understand.}