It is generally agreed upon that FHE and more generally, PETs, for ML/AI is going to be pretty slow. Despite that, there have been many libraries and attempts over the past decade to make these technologies more practical.
Some of the biggest libraries in this area are [TF-Encrypted](https://github.com/tf-encrypted/tf-encrypted) and [Concrete-ML](https://github.com/zama-ai/concrete-ml). I'll notably mention [SPU](https://github.com/secretflow/spu) and [Flower.ai](http://Flower.ai) as well.
Considering that most of these codebases are simply in maintenance mode with the exception of [Flower.ai](http://Flower.ai), what codebases, libraries, etc are considered to be state of the art for FHE-enabled ML/AI in 2026?
Papers are helpful but generally they don't come with codebases and if they do, they are optimized simply for the paper and not real work loads or production usage.
