MiniMax-H3-FastH3-VSA-DataFree-MLX-BF16

MiniMax-H3 FastH3 VSA DataFree MLX BF16

This repository is a self-contained inference package for FastVideo's FastH3 4-step VSA DataFree release on Apple Silicon. It combines the compact MiniMax-H3 FL2VA BF16 runtime package, the pinned FastH3 student adapter, its source-bound AdaLN cache, and the files needed to generate synchronized video and audio with mere.run.

The managed runtime payload is 82.32 GB (82,316,599,044 bytes). It is smaller than the approximately 148 GB upstream checkpoint because it uses inference-only representations:

• The active 50-layer H3 denoising core remains BF16.
• The 50-layer Qwen3-VL conditioner uses MLX affine Q8/group-64. Unused language layers, the LM head, and the final language norm are omitted.
• The video VAE uses FP16. The audio VAE remains FP32 with weight normalization folded into equivalent convolution weights.
• Source-bound inference tables replace the 13B schedule-only AdaLN parameters.
• The FastH3 adapter and its exact four-evaluation AdaLN table are included.

These omissions do not require another download for the supported FastH3 text-to-video workflow. This repository is not a byte-for-byte mirror of the upstream training or Diffusers layout and is not intended for training, arbitrary schedules, or use as a standalone complete Qwen language model.

Use with mere.run

Use a mere.run build that lists video-minimax-h3-fasth3-vsa-datafree-mlx in mere.run model list.

mere.run model pull video-minimax-h3-fasth3-vsa-datafree-mlx \
  --accept-model-license

mere.run video generate \
  "A lighthouse in a winter storm, waves striking the rocks, synchronized wind and surf" \
  --model video-minimax-h3-fasth3-vsa-datafree-mlx \
  --output lighthouse.mp4

After the model pull completes, generation is offline and doesn't fetch a separate base model, text encoder, adapter, VAE, tokenizer, or cache. FastH3 uses four denoising evaluations at released base sigma points 0.999, 0.749, 0.5, and 0.25, followed by the clean endpoint. The runtime fixes adapter strength at 1.0 and uses the released text-only VSA-H3 recipe.

The minimum supported frame count is 22. Frame counts follow 17*n+5, and width and height must be multiples of 32. MiniMax-H3 emits 24 fps video with 32 kHz stereo audio.

Package contents

The repository also contains the tokenizer, base production cache pack, configuration, license, notices, modification disclosure, manifests, conversion receipts, and SHA-256 inventory.

Sources and provenance

• Base model: MiniMaxAI/MiniMax-H3 at ec19cc6daf5d8add9417c18e86b6b58cc6c55027.
• Compact BF16 base: Sawfwair/MiniMax-H3-FL2VA-MLX-BF16 at 6f2c1edb4d31d9110d4a51457ba1d6401a05dfd0.
• FastH3 student transformer: FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree at b65818d41939b5085451074fe8ca8b799f8d4921.
• FastH3 VSA DataFree adapter: FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA at bcf40ca6f457ed66f8badf13514943e390205fca.
• FastVideo reference code: hao-ai-lab/FastVideo at 48a047c05ff4138f20cfa33351499c6ec5945f5d.

SOURCE_MANIFEST.json and transformer.conversion.json describe the compact base. FASTH3_SOURCE_MANIFEST.json and FASTH3_CONVERSION.json describe the student adapter and AdaLN sidecar. SHA256SUMS covers every runtime and provenance file in this package.

Qualification

On August 29, 2026, the package completed a 384 by 256 pixel, 22-frame generation on an Apple M4 Max with 128 GB unified memory. The command used the managed model ID without an adapter argument. End-to-end wall time was 96.37 seconds; the first denoising evaluation, including Metal compilation, took 43.511 seconds, and the next three took 3.104, 3.099, and 3.121 seconds. MLX reported a 42.15 GiB peak, and /usr/bin/time -l reported a 56,740,766,160-byte peak process footprint and 0 process swaps. The host already had system-wide swap, so this does not establish a no-swap system result.

The resulting 0.917-second MP4 contains 22 H.264 frames at 384 by 256 and 24 fps plus 32 kHz stereo AAC audio. Its SHA-256 is e3a86b5a18a94944620dbb3b40bad7cac3689935907c5bf48f8bee181899aabc.

This is bounded runtime qualification, not a general quality claim or proof of equivalence to the full upstream checkpoint.

License and territorial restriction

The model weights are governed by the MiniMax-H3 Community License in LICENSE, not by the mere.run source-code license. At the pinned source revision, the license excludes use, distribution, and display in the United States, European Union, United Kingdom, and Republic of Korea. It also imposes notice, modification-disclosure, attribution, safeguard, and downstream distribution obligations.

This repository is public and ungated, matching the official MiniMaxAI/MiniMax-H3 repository's access posture. Public access does not waive or replace the license. You are responsible for determining whether you may download, use, or redistribute these files and for passing the license and required notices to downstream recipients.

See NOTICE, MODIFICATIONS.md, and the complete LICENSE before download or use.

Limitations

• Supports text-only FastH3 FL2VA generation. It does not support Ref2VA, keyframes, continuation, or arbitrary H3 adapters in this managed profile.
• The VSA path keeps every prefix key tile and the highest-scoring 10% of video key tiles for each video query tile. Prefix queries remain dense.
• The package is designed for Apple Silicon with 128 GB unified memory. Smaller systems are not qualified.
• Outputs can contain artifacts, unsafe content, or inaccurate representations. Review generated media before use or publication.