Originally created to harness the power of Apple silicon and metal, Apple's MLX machine learning framework is taking a major step into diversity thanks to the newly added support of Nvidia GPUs through CUDA. This development is an important milestone, and once developers and researchers open the door to code and test MLX models on MAC hardware, they easily deploy them to high-performance NVIDIA-based GPU clusters.
The integration of CUDA into MLX was the brainchild of developer @ZcBenz, who began experimenting with CUDA support on GitHub a few months ago. These early prototypes have since been refined and phasedly integrated into the main branch of MLX, with basic operations such as matrix multiplication, softmaxing, reduction, sorting, indexing, and more working and validating. Although still ongoing, the progress of the project already shows a transformative change for MLX users.
To understand the impact, consider that Cuda (Nvidia's Compute Unified Device Architecture) works as Metal does with Apple hardware. This is the decisive standard across the machine learning spectrum, from academia to businesses, and promotes dominant frameworks such as Pytorch and Tensorflow with efficient GPU acceleration. By providing CUDA compatibility, MLX now functions as a powerful bridge. It allows developers to create and test models locally in MAC, allowing seamless migration to a CUDA-driven environment for full-scale training and deployment.
Despite this exciting advancement, restrictions remain. The CUDA backend is still under development and some MLX operators are not available yet. Additionally, support for AMD GPUs is also on the roadmap, but it hasn't come to fruition yet. Local support for Nvidia GPUs remains out of reach. Developers cannot directly attach external NVIDIA GPUs or utilize them in MAC MLX workflows. The real advantage is that you prepare the model on the device and export it to a larger environment.
This evolving feature addresses important needs. Not all developers and small teams have continuous access to costly NVIDIA hardware. With CUDA support, MLX allows users to run compact scale development on cost-effective Apple silicon Macs and deploy trained models on more affordable NVIDIA hardware in the cloud or data center. The result is a powerful, cost-effective feedback loop that increases flexibility, experimentation and innovation.
Supporting this initiative is collaboration that suggests industry-level support. Details have not been officially confirmed by Apple, but the project description states that the work is “sponsored by Apple,” suggesting internal approval. The backend prototype began in March 2025 and already shows early signs of progress during testing with CUDA 11.6 on Ubuntu 22.04. A completely refined release is still pending, but the combination of practical kitchen development and corporate support is working well for the final production preparation.
This development has also had a broad impact on Apple's continued investment in machine learning. Last year, the company opened sourced Recurrent Drafter (Redrafter), a speculative deciphering method for LLM inference. The collaboration exemplifies Apple's growing engagement with the GPU ecosystem beyond its own platform, strengthening its strategy towards openness and cross-platform compatibility.
Apple's unified memory architecture and powerful silicon GPUs make MLX a compelling option for machine learning workflows in devices. However, up until now, models that surpass Apple hardware have limited portability. By adopting CUDA, MLX is converted into a middle layer that can target NVIDIA GPUs.
Early adopters have already weighed it. In Github and Hacker news, enthusiastic comments highlight the importance of blending Apple's unified memory capabilities with Nvidia's Cuda ecosystem. One commenter from Hacker News concisely summed up, “I want to connect my Nvidia GPU to an Apple Silicon and allow me to run Cuda.” Not everyone is racing to replace Pytorch or Tensorflow, but many view the integration of MLX and CUDA as a promising addition to existing toolkits.
R/Apple Reddit users noted that while Apple still trains AI models on other platforms (such as Google's TPU), CUDA compatibility creates a new pathway for ML development beyond Cupertino's barrier. Enthusiasts also point to the synergy between Apple's low-latency unified memory setup and Nvidia's performance in large distributed environments.
What's next for MLX? We adopted a test pipeline that was continuously expanded to operator support, tightened documentation, and ultimately CUDA-enabled Darwin (MACOS). Future support for AMD GPUs extends the inclusiveness of the framework in a non-uniform computing environment. Also, as CUDA backends mature, MLX could emerge as connective tissue that makes Apple's development model truly portable. Transform MLX into the mainstay of multi-platform development.
Apple's decision to open source and enrich MLX with CUDA compatibility highlights its commitment to broader machine learning innovation. By enabling a seamless transition from local MAC development to large GPU clusters, MLX becomes a powerful asset for both researchers, startups and enterprise teams. Rather than locking users into one hardware ecosystem, Apple offers flexible and future proof bridges. This could be a strategic pivot of Apple's machine learning playbook, as well as a technical footnote.
