water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8

MiniMax-H3 MLX Argus Calibrated INT8 DiT

MLX-native, activation-aware INT8 DiT/transformer only for MiniMax-H3. Text encoder and VAEs are not included or quantized by this repository.

Format

• Weight-only affine INT8, group size 32
• 254 mlx.nn.QuantizedLinear layers with packed uint32 weights
• Four sensitivity-selected projections retained in BF16
• Eight sharded safetensors, 37,797,142,653 bytes including metadata/license
• AdaLN-drop resident DiT estimate: 22.928 GB before runtime caches
• Source revision: 939557dc319dd91227e30195a763f272ba7f8765
• Runtime implementation commit: d5c1e7995e54bbc99cd8b5e7dbca7fb41d119a39

The optimizer uses real, disjoint calibration/holdout activations and a diagonal-Hessian approximation to ||WX-Q(W)X||². Three weighted affine refits are evaluated per group; groups that do not improve over native MLX round-to-nearest retain the native result.

Measured build gates

• 258/258 activation paths captured and finite
• Calibration objective: 4,572,725.8326 (plain group-32 RTN) → 2,366,819.8175 (candidate), a 48.24% reduction
• Strict MLX reload passed
• 254 INT8 + 4 BF16 exception layers accounted for
• Real saved-weight projection produced finite output
• Real streamed INT8 + BF16 Larry Turbo adapter smoke produced finite refiner, AdaLN-cache, block and final-layer outputs

This is a quality candidate, not a claim of mathematically lossless quantization. A complete 258-layer held-out comparison was not finished before packaging; do not represent it as absolute losslessness.

Usage

Use with the matching Argus-AiTeam/minimax-h3-mac runtime and the separately downloaded MiniMax-H3 pipeline components:

python scripts/generate.py "A cinematic fox running through a misty forest" \
  --checkpoint /path/to/MiniMax-H3/FL2VA \
  --transformer /path/to/MiniMax-H3-MLX-Argus-Calibrated-INT8 \
  --text-encoder /path/to/your/MLX-text-encoder \
  --low-memory --stream-blocks --steps 21 \
  --output out.mp4

--steps counts sigma points in this runtime, so --steps 21 means 20 denoiser evaluations. With the separate Larry v4 Turbo MLX adapter, use --steps 7 to --steps 9 (6–8 evaluations).

License

The weights are derived from MiniMax-H3 and are distributed under the included MiniMax H3 Community License Agreement. Review and accept it before use or redistribution. "Powered by MiniMax H3."