Flox, the Nix Foundation, and NVIDIA Partner for CUDA (flox.dev)

Flox, the Nix Foundation, and NVIDIA have partnered to streamline the building and deployment of CUDA-accelerated software stacks by enabling access to precompiled CUDA packages through the Flox Catalog. This collaboration solves a long-standing bottleneck for AI/ML teams who previously had to endure lengthy source builds for complex CUDA dependencies like TensorRT, PyTorch, and OpenCV. By leveraging Nix’s deterministic package management and Flox’s user-friendly interface, developers can now install multiple versions of CUDA toolkits and libraries side-by-side without containerization or system-wide changes, drastically improving portability, reproducibility, and version management. The key technical breakthrough lies in Flox’s ability to legally redistribute NVIDIA’s proprietary CUDA binaries prebuilt from canonical Nixpkgs recipes, overcoming licensing hurdles that prevented such redistributions in the Nix community. Users of Flox or Nix with Flox’s binary cache can instantly access and deploy curated CUDA stacks that ensure consistent behavior across development, CI, and production environments. While NVIDIA’s official drivers still need to be installed separately, this integration accelerates workflows by minimizing build times from days to minutes and enhances stability by isolating CUDA dependencies in deterministic environments. This partnership unlocks a new era of CUDA development where AI/ML teams benefit from seamless coexistence of multiple CUDA versions, simplified CI maintenance, and reproducibility guarantees critical for production reliability. Early access is available now, with ongoing community feedback shaping future refinements, marking a significant step toward reproducible, portable GPU-accelerated software delivery.