Member-only story
NVIDIA Just Made Rust GPU Programming More Serious. CUDA Oxide 0.3 Is Why
Stable Rust support, experimental CUTLASS CuTe bindings, and real Qwen model execution show how NVIDIA is building a new path for Rust developers into GPU programming.
For years, Rust developers have had a frustrating choice when working close to GPU hardware: write the application in Rust, then cross into CUDA C++ when the performance-critical GPU code begins.
That boundary is starting to look different.
On October 31, 2026, NVIDIA’s CUDA Rust project released CUDA Oxide 0.3, a substantial update that moves the project onto stable Rust 1.99.1, brings its host runtime together with cuTile Rust, and introduces an experimental Rust implementation of CUTLASS CuTe.
The release also demonstrates something more interesting than compiling a small GPU example: cute-rs can run Qwen3-8B using BF16 and NVFP4, as well as Qwen3-TTS, on NVIDIA Blackwell GPUs.
This is not simply another Rust library wrapping an existing CUDA API.
NVIDIA is working toward making Rust a language for writing GPU kernels themselves.
