Native Apple-Silicon package for MiniMax-H3's omni-reference Ref2VA mode.
It accepts ordered image, video and audio references for character, object,
scene, motion, voice and style consistency, and jointly generates video plus
stereo audio.
The DiT and Qwen language trunk use MLX affine Q4 (group size 64).
The Qwen vision tower remains BF16, as do the precision-sensitive projections
and both VAEs. The package is self-contained and is consumed by the Ref2VA
runtime in
mlx-serve.
Limits: up to 9 images, 3 videos, 3 audio clips and 12 files total. Audio
cannot be the only reference. Video/audio references are 2–15 seconds each,
with at most 15 seconds total per modality.
