FastMetal-14B-QAD

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

FastMetal is a family of DMD2-distilled Wan video models with quantization-aware-trained INT8 DiTs. We ship the DiT already quantized, so there is no startup quantization — download, load, generate.

What's inside

This repo is fully self-contained for generation:

Path	Contents
mlx_dit.safetensors / mlx_dit.json	INT8 (affine, group-64) DiT — student checkpoint
ema/	EMA-smoothed variant of the same DiT
text_encoder/, vae/, tokenizer/, scheduler/	everything needed to run standalone (fp16 UMT5 text encoder)

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

# student checkpoint
python examples/inference/basic/mlx_wan_prompt_to_video.py \
  --model-root ./FastMetal-14B-QAD \
  --mlx-checkpoint ./FastMetal-14B-QAD \
  --prompt "a cinematic slow pan over a mountain river at golden hour"

# EMA-smoothed variant
python examples/inference/basic/mlx_wan_prompt_to_video.py \
  --model-root ./FastMetal-14B-QAD \
  --mlx-checkpoint ./FastMetal-14B-QAD/ema \
  --prompt "a cinematic slow pan over a mountain river at golden hour"

Model details

Base model	FastWan 2.1 T2V 14B
Distillation	DMD2, 3 denoising steps
Quantization	affine INT8, group size 64, QAT-trained
Resolution	448×832 (480p), 77 frames
Flow shift	8.0
DiT weights	~15 GB (INT8)

Training

DMD2 distillation of the FastWan 2.1 T2V 14B teacher onto an INT8 student on NVIDIA GB200 clusters, with quantization-aware training (affine INT8, group 64) so the deployed model matches the training objective. Training corpus: FastVideo/Wan-Syn_77x448x832_600k.

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