NVIDIA CUDA Oxide 0.3 brings stable Rust support, CUTLASS CuTe port, Blackwell optimizations

NVIDIA software engineers released CUDA Oxide 0.3, the latest version of their experimental Rust-to-CUDA compiler that has seen rapid development throughout 2026. The release marks a significant maturity milestone by moving from nightly Rust builds to the official stable Rust toolchain, currently version 1.99.1, which broadens accessibility for developers who prefer or require stable compiler guarantees.

The 0.3 release also unifies the host runtime with cutile-rs inside the cuda-rust monorepo, ships an experimental Rust implementation of CUTLASS CuTe called cute-rs, and adds full cuda-gdb DWARF debugging support alongside in-kernel event tracing and register diagnostics. Performance optimizations target both Blackwell and Hopper GPU architectures, with cute-rs validated running Qwen3-8B at BF16 and NVFP4 precision as well as Qwen3-TTS on Blackwell hardware.

What changes in CUDA Oxide 0.3

• Stable Rust toolchain: cuda-oxide now compiles with official stable Rust 1.99.1, eliminating the previous requirement for nightly builds and improving reproducibility for downstream packaging and CI systems.
• Monorepo restructuring: The host runtime has been unified with cutile-rs within the cuda-rust monorepo, consolidating previously separate components into a single repository structure.
• CUTLASS CuTe port: An experimental cute-rs crate provides a Rust interface to CUTLASS CuTe tensor operations, demonstrated running end-to-end Qwen3-8B inference at BF16 + NVFP4 and Qwen3-TTS on NVIDIA Blackwell GPUs.
• Debugging infrastructure: Full cuda-gdb DWARF debugging support is now available, complemented by in-kernel event tracing, register and spill diagnostics, and expanded test coverage.
• Architecture-specific optimizations: The compiler and runtime include performance improvements targeting both Blackwell and Hopper GPU generations.

Why it matters

Moving CUDA Oxide to stable Rust removes a significant barrier to adoption for systems programmers and Linux distributions that cannot easily ship nightly toolchain dependencies, bringing NVIDIA's GPU compute stack closer to standard Rust packaging workflows. The cute-rs CUTLASS port signals NVIDIA's investment in making its highest-performance tensor primitives accessible from Rust, which could accelerate Rust adoption in AI inference workloads that currently dominate Python and C++ ecosystems.