NF4 weight-only quantized Wan2.1-T2V-14B-Diffusers, saved as a complete WanPipeline.

All 40 blocks of the WanTransformer3DModel are quantized to nf4_weightonly — 4-bit NormalFloat weights with a rank-32 low-rank error correction term fitted over 8 iterations, which recovers most of the quality NF4 alone gives up. Activations stay in fp16.

The UMT5 text encoder is not quantized (it is kept in fp16); the VAE is loaded in bf16 and the scheduler is swapped to UniPCMultistepScheduler. Quantization was done with WaveSpeed's xelerate.ao.quantize. Weights are pickled .bin shards, so loading requires use_safetensors=False.

Running the NF4 blocks requires WaveSpeed's xelerate runtime; the low-rank correction term is not part of the stock bitsandbytes NF4 format.

Apache-2.0, inherited from Wan2.1.