Zant – An open-source Zig SDK for neural network deployment on microcontrollers

Zant is an open-source, cross-platform SDK written in Zig and designed to simplify deploying Neural Networks (NN) on microcontrollers. It comprises a suite of tools to import, optimize, and deploy NNs to low-end hardware.

The developers behind the project developed Zant (formerly known as Zig-ant) after noticing many microcontrollers lacked robust deep learning libraries, and made sure it would be on various platforms such as ARM Cortex-M or RISC-V microcontrollers, or even x86 targets. Contrary to platforms like Edge Impulse that focus on network creation, Zant is about deployment and outputs a static, highly optimized library ready to be integrated into any existing work stack.

Zant highlights:

• Optimized Performance – Supports quantization, pruning, and hardware acceleration techniques such as SIMD and GPU offloading.
• Low memory footprint – Zant employs memory pooling, static allocation, and buffer optimization to work on resources-constrained targets.
• Ease of Integration: With a modular design, clear APIs, and comprehensive examples/documentation, integrating Zant into your projects is straightforward.

Zant is suitable for AI-powered computer vision applications (like object detection or optical inspection), predictive maintenance, and adding AI capabilities to autonomous systems like drones and robots. The Zant SDK is still a work-in-progress, but we can find some code and a schedule on GitHub.

Right now, Zant should be able to run MNIST inference on a Raspberry Pi Pico 2 relying on imported models from ONNX, and I assume, working on either Arm or RISC-V cores (TBC) based on the cross-platform aim of the project. Next up is the plan to have YOLO efficiently on Raspberry Pi Pico 2 by the end of April.

Other items on the todo list include implementing a shape tracker to optimize tensor operations, developing a frontend interface for easier interaction with the library, optimizing code generation, expanding ONNX support, and supporting additional microcontrollers and architectures besides just Raspberry Pi RP2350. Based on the schedule, the Zigant library might achieve those goals by the end of Q3 2025.

It’s far from the first time machine learning has been implemented on microcontrollers. As noted above, Edge Impulse is an example, but we also covered BitNetMCU for CH32V003 RISC-V MCU, TinyMax for Arduino-compatible microcontrollers, and TensorFlow Lite for MCUs. The way I understand it, the main benefit of Zant will be the ability to easily deploy a neural network model on heterogeneous platforms with minimal or no modifications to the code.

Via Reddit

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress