MLX Gains CUDA Backend, Bringing Apple's ML Framework to Nvidia GPUs on Linux

Apple's MLX machine learning framework has shipped a CUDA backend in version 0.31.1, extending the framework beyond Apple Silicon for the first time. Installable via pip, it targets Nvidia GPUs with compute capability SM 7.5 or higher, requiring driver 550+ and CUDA 12+, on Linux only. A CPU-only Linux variant ships alongside it, giving MLX three deployment targets: Apple Silicon via Metal, Nvidia GPU via CUDA on Linux, and CPU-only Linux.

The core engineering problem was memory. MLX was built around Apple's unified memory architecture, where CPU and GPU share the same physical pool — a model with no equivalent on discrete Nvidia hardware. Adding CUDA support meant abstracting over fundamentally different memory architectures while keeping MLX's unified API intact. According to the 0.31.1 release documentation, the full feature set carries over: lazy evaluation, function transforms, compilation, distributed communication, and complete LLM inference and training pipelines.

For agent developers, the expansion matters most at the deployment layer. MLX is already a common substrate for <a href="/news/2026-03-14-opentoys-open-source-ai-toy-platform-esp32-voice-cloning">local LLM inference on Mac hardware</a>; opening it to Linux GPU servers increases its addressable base substantially. Existing support for data and tensor parallelism on CUDA devices means it could back multi-GPU agent deployments without pulling in PyTorch or vLLM. A single framework covering both a developer's MacBook and a Linux inference server is now a real option rather than a roadmap item.

MLX now competes directly with PyTorch and JAX across the hardware it covers. PyTorch's dominance runs deep — production pipelines, custom ops, and deployment tooling built over years — and a pip install won't shift that. What the CUDA backend does is eliminate the argument that MLX is a Mac-only convenience layer. The practical test is whether Apple sustains the engineering investment needed to keep the CUDA backend's feature set current as PyTorch's Linux capabilities continue to advance.