NVIDIA has released CUDA 13.2, as detailed on the official CUDA Toolkit page [1].

NVIDIA has released CUDA 13.2, as detailed on the official CUDA Toolkit page [1]. The release extends tile programming support to Ampere and Ada architectures, introduces closures and recursion in cuTile Python, and unifies the ARM ecosystem into a single toolkit. These enhancements simplify GPU development and broaden accessibility. Concurrently, Hugging Face Transformers v5.11.0 [2] adds support for new models (DiffusionGemma, DeepSeek-V3.2) and introduces Triton-based fine-grained fp8/fp4 kernel support, indicating ongoing ecosystem adoption of CUDA-compatible frameworks. While the CUDA updates are incremental, they strengthen NVIDIA's developer platform and may counter emerging alternatives. A Manifold prediction market [3] shows 56.95% confidence that CUDA remains a monopoly through 2027, reflecting some uncertainty but no immediate competitive threat. The combination of active development and ecosystem integration points to sustained relevance, albeit with low confidence due to routine nature of the release and single official source.