// the take
Nx is Elixir's answer to a tensor library with autodiff, and this monorepo bundles it with two swappable backends: EXLA (Google's XLA, for GPU/TPU) and Torchx (LibTorch). It's the base layer under Axon and Bumblebee, so it's for Elixir developers doing ML/numerical work who don't want to shell out to Python for everything.
The backend abstraction actually works — you write the same defn code and swap Binary/EXLA/Torchx underneath, and there's a backend_documentation_test pattern that runs doctests against every backend to catch drift. Linear algebra is implemented properly (QR, SVD, LU, Cholesky live under nx/lib/nx/lin_alg, not just FFI passthroughs). JIT compilation through EXLA to XLA HLO is a real compiler pipeline (see exla/lib/exla/mlir), not a toy wrapper around a Python call.
You're one of maybe a few thousand people running numerical code on the BEAM, so when you hit a missing op or a CUDA edge case, there's no Stack Overflow answer waiting — you're reading exla_cuda.cc yourself. Native builds are heavy: c_src has hand-rolled NIFs, IPC bridging, and separate CUDA/ROCm Dockerfiles, so first-time setup on anything but a supported combo is friction. Three mix projects (nx, exla, torchx) live in one repo with separate mix.exs/README/CHANGELOG each, and the README says they'll eventually split — so tooling and doc links may not be stable long-term. This is tensors + autodiff only; you need Axon for actual neural net layers and Bumblebee for pretrained models, so plan on pulling in siblings, not just this repo.
We read the README, the directory tree and a key source file in September 2026. Nobody pays to be listed. How we judge →