ComfyUI on a Mac: what works on Apple silicon
ComfyUI runs on Apple silicon Macs, and a Mac with a lot of memory can load big models. Speed is the catch, most of all for video.
Updated October 2026
The short answer
- Apple silicon (M1 to M5): ComfyUI runs, using Apple's Metal graphics through PyTorch.
- Memory: a Mac shares one pool of memory between the processor and graphics. A Mac with 64 GB can load models that don't fit on a 24 GB gaming card.
- Speed: several times slower than a current NVIDIA card. Pictures can take a minute or more where an NVIDIA card takes seconds, and video is slow enough that most people don't bother.
- Intel Macs: processor only, which is too slow for current models.
Installing it on a Mac
The ComfyUI desktop installer has a version for Apple silicon Macs that sets up Python and PyTorch for you. You can also install it by hand with Python and Git. Either way, model files go in the same folders as on Windows (see Running ComfyUI locally).
How much memory your Mac needs
On a Mac, the memory you chose when you bought it is also your graphics memory. macOS keeps part of it for itself, so plan on the graphics side getting most of it, not all.
| Mac memory | What runs |
|---|---|
| 8 GB | Stable Diffusion 1.5 at small sizes. Little room for anything newer. |
| 16 GB | SDXL-class image models. Compressed (GGUF) versions of bigger ones, slowly. |
| 24 to 36 GB | Flux-sized image models (about 12 billion parameters) as compressed files. |
| 48 to 64 GB | Most current image models at full precision. Small video models. |
| 96 GB and up | The large image and video models load, but a video clip can take a very long time. |
What causes trouble on a Mac
- fp8 files. Many models are shared as fp8 files to save memory on NVIDIA cards. Apple's graphics backend often can't load them, and the error mentions MPS or Float8. Use the fp16, bf16 or GGUF version.
- NVIDIA-only add-ons. Add-ons built on CUDA, xformers or SageAttention only run on NVIDIA cards. Workflows that depend on them won't run on a Mac.
- Heat. A MacBook Air has no fan and slows down during long jobs. MacBook Pros hold their speed longer but get loud.
- Video. Video models do far more work per job than image models, and on a Mac that turns into waits of tens of minutes or more for a few seconds of clip.
The other route on a Mac
Because most of the AI world runs on NVIDIA cards, many Mac users rent one in the cloud and keep using their Mac for everything else. The options are in Running ComfyUI without a good GPU.
CozyGPU is a Mac app (Apple silicon and Intel) that does this for you: it runs ComfyUI on NVIDIA cards it rents by the second in your own RunPod account, so every model runs at the same speed on any Mac. See the download page.
Common questions
Can I run ComfyUI on a Mac?
Yes, on an Apple silicon Mac (M1 to M5), using Apple's Metal graphics. It runs several times slower than a current NVIDIA graphics card, and some add-ons and file formats only work on NVIDIA. On an Intel Mac it runs on the processor only, which is too slow for current models.
Can a MacBook Air run ComfyUI?
It can run small image models such as SDXL, slowly. With 8 to 16 GB of memory and no fan, it slows down further on long jobs. Current large image models and video aren't practical on it.
Is a Mac or a PC with an NVIDIA card better for ComfyUI?
A PC with an NVIDIA card is faster and better supported. A Mac with a lot of memory can load bigger models than most gaming cards can hold, but it runs them more slowly.
Can I run video models like Wan 2.2 or LTX on a Mac?
On an Apple silicon Mac with a lot of memory, technically yes. A few seconds of video can take tens of minutes or longer, so most Mac users rent a GPU for video instead.
Does ComfyUI work on an Intel Mac?
Only on the processor, which is too slow for current image and video models. Renting a GPU in the cloud is the practical route on an Intel Mac. CozyGPU has a separate download for Intel Macs, and it runs just as fast there because the models run on cloud GPUs.
What does an MPS or Float8 error on a Mac mean?
Often that the model file is in fp8 format, which Apple's graphics backend for PyTorch (MPS) doesn't handle. Use the fp16, bf16 or GGUF version of the same model instead.
How much memory does a Mac need for ComfyUI?
16 GB for SDXL-class pictures, 32 GB or more for current image models as compressed files, and 64 GB or more for full-size image models. macOS keeps part of the memory for itself, so the graphics side can use most of it, not all.
CozyGPU rents the graphics card for you, sets 12 open-source models up on it and gives you a simple app on your Mac or PC. It runs ComfyUI for you on a GPU in your own RunPod account, billed by the second, so nothing is charged while you're not using it. $20 a month, plus the GPU time.