FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4

FastH3 Preview v1 Dense DataFree MLX INT4

This repository contains the ready-to-run INT4 MLX DiT for FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree. It avoids the local conversion step and its temporary unified-memory requirement on Apple silicon.

The checkpoint is the dense-attention, step-1000, four-forward Preview v1 model. Quantization is affine, weight-only INT4 with group size 64. Activations remain BF16. This export is dense-only and does not support --vsa.

Files

File	Purpose
mlx_h3_dit.safetensors	Quantized H3 DiT weights, 10.74 GiB
mlx_h3_dit.json	Architecture, quantization, and cached AdaLN schedule
conversion_manifest.json	Source revision, converter version, validation, and checksums

Weights SHA-256: dcee86c17197c621a0e3d3ba9541cb6ef5764cd0c258beb7b044fb12b217e948

Download

Download the shared H3 components once. They include the tokenizer, Qwen3-VL text encoder, video VAE, and audio VAE:

hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree \
  --revision f624f08c6c279ab43534c003e556fc5b295b6558 \
  --local-dir ./FastH3-Preview-v1-Dense-DataFree

Then download this ready-to-run MLX DiT:

hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree-MLX-INT4 \
  --local-dir ./FastH3-MLX-INT4

Run on Apple silicon

From a current FastVideo checkout:

uv venv --python 3.12 --seed
source .venv/bin/activate
uv pip install -e ".[mlx]"

python examples/inference/basic/mlx_fasth3.py \
  --model-root ./FastH3-Preview-v1-Dense-DataFree \
  --mlx-checkpoint ./FastH3-MLX-INT4 \
  --prompt '(S1) A presenter says <d>[English] Fast H3 runs on Apple silicon.</d>' \
  --height 480 --width 832 --num-frames 124 --steps 4 --seed 2026 \
  --output-path ./outputs/fasth3_int4.mp4

The MLX runtime loads one heavyweight component at a time. The source model's MiniMax H3 Community License applies to this converted checkpoint.

Validation and provenance

The conversion used FastVideo commit cf6a00b9be4675602126d6aeab902ad9a74810ea with MLX 0.32.2 on an Apple M4 Max. FastVideo verified all 13 source transformer shards, opened the safetensors header, validated 1,464 tensors and the quantization manifest, loaded the checkpoint, and completed a 124-frame 832x480 generation with the full H3 VAE.