FastVideo FastH3-8-Step-V2 (MLX INT8 Quantized with VSA)

This repository contains the MLX-native INT8 quantized checkpoint of FastVideo/FastVideo-FastH3-8-Step-V2, fully optimized for Apple Silicon Macs (M2/M3/M4 Max/Ultra).

Key Highlights

Format: MLX affine quantized INT8 (group_size=64, bits=8) with VSA (Variable Sparse Attention) support.
Size: ~22 GB safetensors (reduced from original 65+ GB BF16).
Quality: Highest-fidelity quantized version, preserving virtually 100% of the original BF16 visual and generative capability.
Performance: End-to-end inference takes ~31 minutes on Apple M3 Max (128GB) with high-quality tiled VAE decoding (8 denoising steps).

How to Use

Method 1: Python API via FastVideo

git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
uv sync --extra mlx

from pathlib import Path
from fastvideo.mlx_runtime.minimax_h3_pipeline import MiniMaxH3MLXPipeline

prompt = "(S1) A presenter says <d>[English] FastVideo runs on Apple Silicon.</d>"
model_root = Path("models/FastH3-8-Step-V2")
mlx_checkpoint = Path("models/FastH3-MLX-vsa/int8")

pipeline = MiniMaxH3MLXPipeline(
    model_root=model_root,
    mlx_dit_checkpoint=mlx_checkpoint,
    video_decode_backend="h3-vae",
    vae_dtype="fp16",
)

result = pipeline.generate(
    prompt=prompt,
    output_path="outputs/fasth3_8step_int8.mp4",
    height=480,
    width=832,
    num_frames=124,
    seed=2026,
    num_steps=8,
    tiled_video_decode=True,
    vsa=True,
    vsa_sparsity=0.8,
    vsa_tile_size=64,
    vsa_impl="simd",
)
print("Video saved to:", result.video_path)

Method 2: Command Line Interface (CLI)

python -m fastvideo.mlx_runtime.minimax_h3_pipeline \
    --prompt "(S1) A presenter says <d>[English] FastVideo runs on Apple Silicon.</d>" \
    --model-root models/FastH3-8-Step-V2 \
    --mlx-dit-checkpoint models/FastH3-MLX-vsa/int8 \
    --output outputs/output.mp4 \
    --height 480 \
    --width 832 \
    --num-frames 124 \
    --num-steps 8 \
    --tiled-video-decode \
    --vsa \
    --vsa-sparsity 0.8

Method 3: Direct Hub Download & Local Deployment

# Download using huggingface hf tool
hf download vanch007/FastVideo-FastH3-8-Step-V2-MLX-INT8 --local-dir models/FastH3-MLX-vsa/int8

Hardware Recommendations

Recommended: Apple M2/M3/M4 Max or Ultra with 64GB or 128GB Unified Memory.
Power: Keep power adapter connected and enable "High Power Mode" in macOS Battery Settings for optimal sustained GPU throughput.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized