⚡ Quick Answer
fp8 is a way of storing a model's numbers using 8 bits of precision instead of the usual 16. Think of it like rounding a long decimal to fewer digits — you lose a tiny bit of exactness, but the file gets roughly half the size and needs roughly half the VRAM to load.
Because the quality loss is usually small, fp8 has become a common "middle ground" release alongside a checkpoint's full-precision version.
Where You'll See It
Model download pages often list an "fp8" version next to the full bf16 or fp16 release. Some ComfyUI loader nodes also have a weight_dtype setting where you can pick an fp8 option (like fp8_e4m3fn) even when loading a full-precision file.
Quick Example
A checkpoint's bf16 version might need around 24 GB of VRAM to load. Its fp8 version needs roughly half that, letting it fit comfortably on a 12 GB card with output that looks very close to the original.
Frequently Asked Questions
See It In Action
Not sure which version to download?
Our VRAM & Hardware Guide matches models to what your GPU can actually run.
Published: 2026-09-17 · Last updated: 2026-09-17
Join the discussion
Sign in to leave a comment or reply
No comments yet
Be the first to share your thoughts!
