MiniMax-H3-FastH3-VSA-DataFree-MLX-Q8

MiniMax-H3 FastH3 VSA DataFree MLX Q8

This is a self-contained Apple Silicon inference package for FastVideo's four-evaluation FastH3 VSA DataFree MiniMax-H3 student. It includes the Q8 transformer, Q8 text encoder, both VAEs, tokenizer, VSA compression gates, and all AdaLN tables required by mere.run. Inference does not download another base model, adapter, text encoder, or cache.

License notice

These are modified MiniMax-H3 weights governed by the included MiniMax-H3 Community License Agreement (LICENSE) and notice (NOTICE). The license excludes use, distribution, and display in the United States, European Union, United Kingdom, and Republic of Korea and imposes downstream notice and safety obligations. Downloading this ungated repository does not waive those terms.

Conversion

The finalized FastH3 low-rank updates and direct differences were merged into the compact BF16 student with FP32 accumulation. The 208 inference core linear weights were rounded once to BF16 and encoded as MLX affine Q8/group-64. The 50 VSA compression gates use the same encoding. Cache-covered AdaLN updates remain in the source-bound inference table. See MODIFICATIONS.md, FASTH3_SOURCE_MANIFEST.json, FASTH3_CONVERSION.json, and SHA256SUMS for the complete provenance and integrity receipts.

Run with mere.run

Use a mere.run build that includes premerged FastH3 Q8 support:

mere.run video generate \
  "A lighthouse in a winter storm with synchronized wind and surf." \
  --model-root /path/to/this/package \
  --h3-adapter /path/to/this/package/fastvideo_fasth3_4step_v1_vsa_datafree_rank64.safetensors \
  --width 512 --height 320 --num-frames 22 \
  --output lighthouse.mp4

The public managed model ID is video-minimax-h3-fasth3-vsa-datafree-mlx. It selects the released schedule, unit adapter strength, compiled block execution, 64-token VSA tiles, per-head top-k routing at 90% video-key sparsity, and learned pooled-value compression.