Some third-party vendors (e.g., Anaconda, Canonical) have backported a few security fixes to their CUDA 12.6 packages, but these are . Use with caution.
CUDA 12.6 December introduces . Developers can now define "soft" and "hard" constraints for memory migration. Instead of the driver blindly swapping pages to slow system RAM, developers can now tag specific allocations as "High Priority" (keep on VRAM at all costs) versus "Streamable" (allowed to migrate to CPU RAM or NVMe via GPUDirect Storage transparently). cuda 12.6 update news december 2025
The December update introduces . Previously, changing tensor dimensions often required re-capturing the entire graph or maintaining multiple graph instances. CUDA 12.6 now natively supports dynamic input shapes within a single captured graph, utilizing a new memory pooling backend that automatically resizes without the overhead of graph destruction and recreation. This effectively lowers the barrier for real-time video processing and dynamic batch inference pipelines. Some third-party vendors (e
The CUDA 12.6 toolkit includes several updates to make development easier and more efficient: Developers can now define "soft" and "hard" constraints
Previously, developers saw kernel execution times in isolation. The updated Nsight Systems now integrates with popular front-end frameworks (PyTorch 2.6, JAX) to map GPU kernels directly to the computational graph of the model.
As of December 2025, NVIDIA has not announced or released a CUDA 12.6 update. The last stable production release in the CUDA 12.x series is , followed by CUDA 12.8 (December 2024) and CUDA 12.9 (June 2025).
As the high-performance computing (HPC) and AI communities wind down a landmark year, NVIDIA has delivered a final, substantial gift to the developer ecosystem. The December 2025 update to the CUDA ecosystem (build 12.6.x) has landed, and while it may not carry the marketing flash of a major version number jump, industry insiders are calling it one of the most critical stability and optimization releases in recent memory.