Aidge v0.6: Advancing Embedded AI with Efficiency and Flexibility 

Aidge is designed to help developers create, optimize, and deploy neural networks in environments where resources are limited and performance constraints are high. The platform provides an end-to-end design flow, supports the ONNX standard for interoperability, and offers fine-grained control over code generation and hardware adaptation. 
Version 0.6 introduces several major advancements: 

•  Expanded ONNX support and simplification

 ONNX import and export capabilities have been enhanced, and a new simplification tool has been added to reduce computational complexity by folding constants, fusing batch normalization layers, and removing unnecessary nodes. 

• Advanced optimization techniques

Aidge now supports both Post Training Quantization using bit shifting (for true quantization) and Quantization-Aware Training (QAT) using the LSQ method. Aidge also includes the most powerful tensor compression technique with a simple and automatic rank selection. These techniques allow models to be optimized for deployment without compromising accuracy, particularly useful for running inference on constrained hardware.

• Improved backend performance and capabilities 

Backend modules now support more operators, data types, and scheduling options. OpenMP support has been added to several operators, improving parallel execution on multicore processors. A new benchmark framework makes it easier to compare inference speed across CPU, CUDA, and C++ exports , and even against PyTorch and ONNX Runtime!

• Spiking Neural Networks (SNNs)

This release marks the first implementation of spiking neural networks in Aidge, enabling brain-inspired computation for ultra-low-power applications. New operators such as Heaviside and Leaky Integrate and Fire (LIF) have been added, along with tools to generate spike trains from standard inputs. 

• Developer-centric improvements

The platform continues to prioritize usability. Updated tutorials, documentation, and example scripts support quick adoption. The Python interface has been improved, and cross-module integration is more robust. Aidge remains lightweight and modular, ensuring it can be adapted to a broad range of use cases. 

Aidge is available as open source and is actively maintained on the Eclipse Foundation's GitLab. Each module is versioned and can be used independently, fostering transparency and collaboration. 
The development roadmap is ambitious: version 0.7 is expected in July 2025, followed by version 0.8 in September. Future updates will further refine training, compression, and hardware deployment capabilities. 

As AI becomes ubiquitous in everyday devices, there is a growing need for frameworks that are efficient, flexible, and independent of proprietary ecosystems. Aidge addresses this by offering a fully open, customizable solution for embedded AI, built with European expertise and aligned with the goals of the DeepGreen project. 
For more information or to access the source code, visit the Aidge platform webpage and join the growing community of contributors: https://projects.eclipse.org/projects/technology.aidge