ComfyUI needs Python, a recent PyTorch and a supported GPU (NVIDIA, AMD, Intel, Apple Silicon or Ascend NPU), though it can also run on CPU only, slowly. The README says it can run even the biggest open-source models on as little as 4GB of VRAM plus 8GB of RAM, using asynchronous weight streaming.
From the README and repo
| Requirement | What the source says |
|---|---|
| Python | 3.13 is very well supported; 3.14 works but some custom nodes may have issues; the project's pyproject.toml requires 3.10 or newer |
| PyTorch | 2.7 is minimally supported; the latest major version is strongly recommended |
| NVIDIA CUDA build | A cu130 or newer PyTorch is required on the 20 series and above |
| GPUs | NVIDIA, AMD, Intel, Apple Silicon and Ascend NPUs have install paths |
| CPU only | python main.py --cpu runs everything on the CPU (slow) |
| Windows portable | Ships with Python 3.13 and PyTorch CUDA 13.0 |
Those 4GB and 8GB figures are the floor, not a target: how much memory a given model needs is a separate question, covered in how much VRAM ComfyUI needs.
Install it
See how to install ComfyUI on Linux for the commands. Custom nodes bring their own dependencies on top: the ComfyUI custom node install guide lists the traps for the common ones.
If you would rather not set the machine up yourself, the ComfyUI template on Aquanode provisions the box, installs ComfyUI and starts it when you deploy it; once the deployment is Ready, the Overview tab shows the ComfyUI URL.
Sources
- ComfyUI README: Installing. github.com
- ComfyUI pyproject.toml. github.com
- ComfyUI cli_args.py (--cpu). github.com