noctuashap/MiniMax-H3-pruned-r16

MiniMax-H3, rank-16 AdaLN

This is MiniMax-H3 with the AdaLN modulation projections factorized down to rank 16, applied to both DiT partitions. It uses about 39% fewer parameters and 23 GiB less VRAM than the official weights, and it runs at the same speed.

	official	this
parameters	33.12B	20.17B
peak VRAM / GPU	88.3 GiB	65.5 GiB
latency	95.05 s	95.44 s

Those numbers were measured over 50 steps at 1344×768, 124 frames, on 4×GB200.

Use

Point FastVideo at this repo instead of MiniMaxAI/MiniMax-H3 and nothing else needs to change. T2VA, FL2VA and Ref2VA all work.

python examples/inference/basic/basic_minimax_h3_t2v.py \
    --model-path noctuashap/MiniMax-H3-pruned-r16 --prompt "..."

You will need a build of FastVideo that supports rank-reduced AdaLN.

Quality

The output is equivalent to the official weights, though not bit-identical, because H3 relocates the sample under any numerical change.

	MS-SSIM	audio cosine
T2VA	0.903	0.99693
Ref2VA	0.883	0.99861
two different seeds	0.211	−0.017

The last row is what makes the others readable. Two completely unrelated samples score 0.211, so this checkpoint sits roughly 86% of the way toward identical. These numbers come from one prompt per task, compared against the full-rank output.

Attribution

Derived from MiniMaxAI/MiniMax-H3 and distributed under the MiniMax H3 Community License Agreement. All of the model's capabilities are MiniMax's. The "pruned" naming follows Comfy-Org/MiniMax-H3, which applied a rank-8 version of the same idea.