Flare

Currently only supports non-batch operations, which the README admits is a limitation for efficient training on large datasets, with auto-batching still in development.

Requires installing Intel MKL and, on macOS, disabling System Integrity Protection for Neanderthal to load native libraries, making initial configuration cumbersome compared to mainstream frameworks.

Being a Clojure-specific library with a smaller community, it lacks the extensive tooling, pre-trained models, and frequent updates of larger projects like PyTorch, as hinted by its 2017-era focus.

While Neanderthal supports CUDA and OpenCL, Flare's GPU support is noted as not fully implemented in the README, potentially limiting high-performance GPU usage for complex models.