Local AI tools often ask for a particular CUDA version, and error messages about CUDA are some of the most common setup problems. The confusing part is that a Windows PC can have three different "CUDA versions" at once. This guide is based on NVIDIA's nvidia-smi documentation, CUDA installation guide and CUDA Toolkit release notes, and PyTorch's install page. Forge PC hasn't tested every app's CUDA requirements, so always check the app's own install page too.
The three CUDA versions
What | How to check | What it means |
|---|
Driver's CUDA version | nvidia-smi
| The latest CUDA version your installed NVIDIA driver supports |
CUDA Toolkit version | nvcc -V
| The developer toolkit, only present if you installed it |
An app's CUDA build | In the app, e.g. PyTorch | The CUDA version that build of the app was made for |
You only need the CUDA Toolkit if you compile CUDA software yourself. Apps installed as ready-made builds, such as PyTorch's pip packages, are built for a particular CUDA version, so for most users the driver is the part that has to keep up.
1. Check the driver's CUDA version with nvidia-smi
Open Terminal or Command Prompt and run:
nvidia-smi
The header shows Driver Version and CUDA Version. NVIDIA's documentation is precise about what the second number means: it's the CUDA "User Mode Driver" version, "the latest CUDA version supported by the driver". It says this is usually, but not always, the version of the CUDA Toolkit installed. In other words, it's a ceiling, not proof that a toolkit is installed.
If an app says it needs a newer CUDA version than nvidia-smi shows, update your driver first; see how to update graphics drivers.
If you've installed the CUDA Toolkit to compile software, NVIDIA's Windows installation guide says you can check its version by running:
nvcc -V
in a Command Prompt window. If Windows says nvcc isn't recognised, the toolkit either isn't installed or isn't on your PATH, which is fine if you only use apps that bundle CUDA.
Note that NVIDIA's release notes say the driver is no longer bundled with the CUDA Toolkit on Windows from CUDA 13.1 onwards, so install the driver separately from NVIDIA's driver downloads page.
3. Check that PyTorch can use CUDA
PyTorch is installed in builds for a specific Compute Platform: its Start Locally selector offers CUDA versions (CUDA 12.6 and 12.8 were among the options listed when we checked) or a CPU-only build. The CUDA version you picked there is the one that matters for PyTorch, regardless of whether you installed a CUDA Toolkit. To check PyTorch can actually see your GPU, run this in a Python session inside your environment:
import torch
print(torch.cuda.is_available())
If it prints False, you most likely installed the CPU-only build or your driver is too old. Our guide to installing PyTorch with CUDA on Windows explains how to get the right build.
How the versions need to line up
NVIDIA's release notes say the CUDA driver is backward compatible: applications built with a particular CUDA Toolkit version keep working on later drivers. Each CUDA release also has a corresponding driver branch. NVIDIA's current table lists:
CUDA Toolkit | Driver branch |
|---|
13.5 | R620 |
13.4 | R615 |
13.3 | R610 |
13.2 | R595 |
13.1 | R590 |
13.0 | R580 |
NVIDIA adds that existing CUDA 13.x applications run on drivers 580 or newer under minor version compatibility, but CUDA 13.5's new features need an R620 or later driver. The practical rule: your driver's CUDA version (from nvidia-smi) should be equal to or higher than the CUDA version your app or toolkit needs.
If you run AI tools under Linux on Windows, see how to set up WSL2 for GPU local AI.
Our verdict
Run nvidia-smi first: its CUDA Version is the highest CUDA version your driver supports, and it's the number that matters for most local AI apps. Use nvcc -V only if you've installed the CUDA Toolkit, and torch.cuda.is_available() to confirm PyTorch's CUDA build can see your GPU. If an app needs a newer CUDA version than nvidia-smi reports, update the NVIDIA driver. Forge PC hasn't tested every app's requirements; this guide is based on NVIDIA's documentation.