Wan2.1-T2V-1.3B — Ternary Quantized (tritplane3)

Storage format note. This checkpoint is stored dequantized to FP16. Loaded with stock diffusers it runs like the base model — the memory/compute benefits of ternary are not realized in this format. It demonstrates that ternary PTQ preserves generation quality. The packed 2-bit ternary runtime (ternary-quant) currently targets transformers LLMs/VLMs, not diffusers video pipelines, so there is no accelerated ternary inference path for this model yet. Packed weights and a diffusers runtime are on the roadmap.

First publicly available ternary-quantized Wan video model on HuggingFace.

Ternary-quantized version of Wan-AI/Wan2.1-T2V-1.3B-Diffusers — Alibaba's text-to-video DiT model. Produced with ternary-quant applied to the WanTransformer3DModel.

Specifications

Property	Value
Base Model	Wan-AI/Wan2.1-T2V-1.3B-Diffusers
Architecture	WanTransformer3DModel (DiT)
Transformer Params	1.42B
Quantization	tritplane3 (306 linear layers)
Text Encoder (UMT5-XXL)	FP16 (preserved)
VAE (WanVAE)	FP16 (preserved)
License	Apache 2.0

Size

Method	Transformer Size
FP16 (original)	2.84 GB
Ternary tritplane3 (theoretical packed)	~1.42 GB
In this repo (dequantized FP16)	2.7 GB

Weights have ternary precision but stored as FP16 for drop-in diffusers compatibility.

Usage

import torch
from diffusers import WanPipeline
from diffusers.utils import export_to_video

pipe = WanPipeline.from_pretrained(
    "AsadIsmail/Wan2.1-T2V-1.3B-ternary",
    torch_dtype=torch.bfloat16,
)
pipe.to("mps")  # or "cuda"

output = pipe(
    prompt="a cat walking on green grass",
    num_frames=81,
    num_inference_steps=30,
).frames[0]
export_to_video(output, "output.mp4", fps=16)