NVIDIA cuda-oxide 0.3 builds on stable Rust and runs Qwen3-8B in Rust on Blackwell

NVIDIA published cuda-oxide v0.3.0 on 31 October 2026. cuda-oxide is its compiler for writing CUDA kernels in Rust. The release notes call it the project’s largest so far, with “nearly 600 commits from 45+ contributors since v0.2.1.”

When we covered NVIDIA’s two CUDA Rust tracks in September, cuda-oxide was an early alpha installed with a pinned nightly compiler, while its sibling cutile-rs already ran on stable Rust. This release changes both points.

What changed

• Stable toolchain. The whole stack now builds on stable Rust 1.99.1. It does this by pairing rustup’s rustc-dev component (librustc_driver.so and rustc_public) with the cargo-oxide tool. Those are shipped on the stable channel but are compiler-internal interfaces, so this is “builds on the stable toolchain”, not “uses only stable APIs”.
• One repository, one host runtime. cuda-oxide (SIMT kernels) and cutile-rs (Tile IR kernels) now live together in NVIDIA/cuda-rust on shared host crates (cuda-core, cuda-async, cuda-bindings, v0.5.0). The same DeviceBuffer, Stream, Event and CudaGraph pass between both kinds of kernel.
• CuTe in Rust, as a preview. The experimental cute-rs-preview brings CUTLASS CuTe 4.7 layouts, TMA pipelines and Blackwell tensor-core MMAs into Rust’s type system for sm_100a (B200) and sm_120a (RTX 5090), with 14 standalone kernel examples.
• Two whole models in Rust. The notes list full 36-layer Qwen3-8B inference in BF16 and NVFP4 (NVIDIA says NVFP4 gives a “3.56x smaller memory footprint”) and the Qwen3-TTS 1.7B codec and speech decoder, with “zero Python runtime”.
• Coverage. Over 1,240 device intrinsics checked against PTX ISA floors from sm_75 to sm_100a and sm_120a, plus 120 libm functions lowered to libdevice.

What is missing

The notes contain no benchmark against CUDA C++. NVIDIA’s wording is that codegen work “close[s] the gap with hand-tuned CUDA C++”; that is its description, not a measured result.

Install is from the repository. On crates.io, the cuda-core crate is at 0.4.0 (25 September), not the 0.5.0 in the notes, and the cuda-oxide name there belongs to an unrelated 2021 crate. The CuTe models need Blackwell GPUs.

Rust inference on NVIDIA hardware is a busy corner: Respan’s Span-02 ships its weights for a Rust and CUDA engine only.

Why this matters

The September post’s main blocker for Rust GPU teams was the pinned nightly; that is gone, and one runtime across SIMT and Tile kernels removes a second wrapper layer. Two caveats keep it in the experiment column for now: the build depends on rustc-dev internals that can change between compiler releases, and there are no published performance numbers to weigh against CUDA C++.