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.
| FP8 | GGUF Q4/Q5 | |
|---|---|---|
| Typical GPU | 4090 / 5090 / A100+ | 12–24 GB consumer |
| Graph | Dense FP8 GEMM | Dequant tile → GEMM |
| Closest quality to bf16 | Usually FP8 | Q8 close; Q4 softer |
| Loader | DiffSynth, native Comfy | ComfyUI-GGUF, llama.cpp-style |
| MoE | Keep both experts; maybe offload | Often 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
