FastMetal-5B-QAD

3-step text-to-video, INT8 pre-quantized for Apple Silicon.

The mid-tier FastMetal model — a DMD2-distilled Wan2.2 TI2V 5B with a quantization-aware-trained INT8 DiT. 720p-native, pre-quantized: no startup quantization.

What's inside

Path	Contents
mlx_dit.safetensors / mlx_dit.json	INT8 (affine, group-64) DiT
text_encoder/, vae/, tokenizer/, scheduler/	everything needed to run standalone

Quickstart

Requires macOS with Apple silicon (MPS) and Python 3.11+:

pip install torch transformers mlx safetensors av imageio imageio-ffmpeg
git clone https://github.com/FastVideo/FastVideo.git
cd FastVideo

python examples/inference/basic/mlx_wan22_generate.py \
  --text-encoder-root ./FastMetal-5B-QAD \
  --mlx-checkpoint ./FastMetal-5B-QAD \
  --vae-root ./FastMetal-5B-QAD/vae \
  --prompt "a river winding through a fantasy valley at golden hour" \
  --fast

Model details

Base model	Wan 2.2 TI2V 5B
Distillation	DMD2, 3 denoising steps
Quantization	affine INT8, group size 64, QAT-trained
Resolution	704×1280 (720p), 121 frames
Flow shift	5.0
DiT weights	~5 GB (INT8)

Training

DMD2 distillation of the Wan 2.2 TI2V 5B teacher onto an INT8 student on NVIDIA GB200 clusters, with quantization-aware training (affine INT8, group 64). Training corpus: FastVideo/Wan2.2-Syn-121x704x1280_32k.

FastMetal family

Model	Tier
FastMetal-1.3B-QAD	Entry — 16 GB+ class Macs
FastMetal-5B-QAD	Mid — 720p
FastMetal-14B-QAD	Quality — 24 GB+/ Ideally 36 Macs