Build + benchmark kit for FastH3 Omni on MLX

0ac7396 verified 7 days ago

1.26 kB

license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: FastVideo/FastVideo-FastH3-Omni-8-Step-V1
tags:
  - mlx
  - apple-silicon
  - fastvideo
  - minimax-h3
  - fasth3-omni
  - benchmark

FastH3 Omni on Apple Silicon (MLX) — build + benchmark kit

FastH3 Omni (reference-conditioned video + audio) runs natively on Apple Silicon through FastVideo's MLX runtime (hao-ai-lab/FastVideo#2000).

This repository holds the kit, not the weights:

bench/prepare_model_dir.sh builds the MLX model directory from the public releases (FastVideo/FastVideo-FastH3-Omni-8-Step-V1, FastVideo/FastH3-Omni-NVFP4-QAD, FastVideo/FastH3-NVFP4-VAE-Decoder, lightx2v/LightVAE) at pinned revisions: INT6 (default), INT8 or BF16 transformer, NVFP4 text-encoder cache, FastH3 36L/8L decoders and the LightVAE Lite H3 reference encoder. It verifies every file against SHA-256.
bench/run_matrix.py runs the FastH3 Omni reference workloads and writes a latency table. See bench/README.md.

Licenses: MiniMax-H3 Community License for the model weights; Qwen3-VL and LightVAE are Apache-2.0 (see bench/licenses/).