What it does

This is an awesome-list in the classic GitHub tradition: a hand-maintained index of hardware, software, and reading material for running AI inference on resource-constrained devices. Think microcontrollers, camera modules with built-in NPUs, and single-board computers that sip milliwatts instead of watts.

The interesting bit

The hardware section reads like a taxonomy of ambition. You’ve got Google Coral and NVIDIA Jetson at the “plenty of room” end, but also chips like the Kendryte K210 (RISC-V with 64 “KLUs”) and the Ambiq Apollo3 Blue (Cortex-M4) that are genuinely tiny. The software side is equally sprawling — from TensorFlow Lite (now LiteRT) down to onnx2c, which literally compiles neural networks to plain C for “Tiny ML.”

Key highlights

• Hardware spans the full power curve: SDR transceivers with deep learning (AIR-T), sub-dollar RISC-V chips, and the new Raspberry Pi AI HAT+ with 26 TOPS
• Software includes commercial platforms (Edge Impulse, Qeexo AutoML) alongside bare-metal compilers (nncase, onnx2c, emlearn)
• Notable recent additions: STM32N6 with Cortex-M55 + NPU, Nordic’s nRF54L with an Axon NPU for wireless sensor-type applications, and Eclipse Aidge as an open-source deployment framework
• Curated with explicit rules: no duplicates, short unbiased descriptions, one commit per suggestion
• CC0 licensed — the maintainer explicitly waived all rights

Caveats

• No benchmarks or comparative data; you’ll need to cross-reference against your own power/latency/accuracy constraints
• Some hardware links point to product pages that may go stale (the Movidius link already redirects to Intel’s generic AI page)
• The “Software” section mixes inference frameworks, training tools, and commercial services without clear categorization

Verdict

Grab this if you’re in the research or early prototyping phase of an edge AI project and need to map the landscape before committing to a platform. Skip it if you already know you need TensorFlow Lite on a specific chip — you’ll get more from that project’s own docs.