Back to wiki Inference

FP8 vs GGUF

FP8 and GGUF both shrink Wan, but they are not interchangeable. FP8 is a GPU tensor-core element type on a dense graph. GGUF is a block-quantized file format (Q4–Q8) designed for loaders that dequantize tiles on the fly.

e4m3Q4_KComfyUI
On this page 4
FP8 versus GGUF versus other Wan compression paths
FP8 stays in the CUDA matmul path. GGUF is a storage/runtime format from the llama.cpp world.

FP8 (e4m3 / e5m2)

  • Native on Ada/Hopper/Blackwell tensor cores.
  • Usually one FP8 DiT plus an FP32/BF16 residual path; VAE often stays higher precision.
  • Tools: DiffSynth-Studio, LightX2V, some ComfyUI fp8_e4m3fn safetensors.
  • Quality: typically the closest to bf16 of the cheap options if calibration is sane.

GGUF (Q8_0, Q5_K, Q4_K, …)

  • Block quantization with optional importance matrices (K-quants).
  • Popular in ComfyUI-GGUF for WAN/Flux/Qwen so a 16 GB card can load a 27B-total MoE.
  • Speed depends on the dequant kernel. On a strong GPU, Q4 can be slower than FP8 because of dequant + less tensor-core use.
  • Two experts still mean two GGUF files for A14B in many graphs.
FP8GGUF Q4/Q5
Typical GPU4090 / 5090 / A100+12–24 GB consumer
GraphDense FP8 GEMMDequant tile → GEMM
Closest quality to bf16Usually FP8Q8 close; Q4 softer
LoaderDiffSynth, native ComfyComfyUI-GGUF, llama.cpp-style
MoEKeep both experts; maybe offloadOften split files + sequential offload

FAQ

Is GGUF “the quantized official model”?
It is a community re-pack. Hash and VAE must match the parent (A14B vs 5B vs I2V).
NVFP4?
Newer LightX2V-style 4-bit GPU formats. Closer to FP8 in spirit (GPU GEMM) than to GGUF files.

Primary sources

Checked against Wan-Video/Wan2.2 README · Aug 28, 2026

© 2026 wan2.video