This is the self-contained, inference-ready MiniMax-H3 Ref2VA artifact for
native Swift/MLX inference in mere.run.
It generates synchronized 24 fps RGB video and 32 kHz stereo audio from an
ordered presentation of image, video (including its soundtrack), and audio
references.
The transformer is MLX affine INT8 with group size 64. Eight-bit is the lowest
published Ref2VA precision because lower precision did not meet the visual
quality bar. The Qwen3-VL conditioner is also MLX affine INT8/group-64; the
video VAE is FP16 and the audio VAE is FP32.
The package includes a source-bound 31-point AdaLN cache. mere.run validates it
against the immutable transformer SHA-256, uses the exact released schedule,
and can omit the transformer's schedule-only AdaLN/time-embedding weights at
load time. No post-pull model optimize step is required.
Install
mere.run model pull video-minimax-h3-ref2va-mlx --accept-model-license
The managed download is explicit and never occurs automatically. Review
LICENSE before pulling or using the model.
Generate
Reference order is semantic and is preserved exactly:
bash
1mere.run video generate \2"keep the subject, follow the camera move, and match the vocal rhythm"\3 --model video-minimax-h3-ref2va-mlx \4 --reference image:./subject.png \5 --reference video:./camera-and-soundtrack.mp4 \6 --reference audio:./voice.wav \7 --num-frames 124\8 --output ./referenced-h3.mp4
The released limits are 12 references total: at most 9 images, 3 videos, and
3 audio clips. Audio cannot be the only reference type. Ref2VA regenerates
synchronized audio conditioned by the presentation; it does not copy the
input waveform into the output.
Provenance
The Ref2VA transformer was converted from exactly:
The converter validates each tensor's embedded ConvRot metadata. The source
uses group 256 for 200 transformer matrices and group 64 for 50 AdaLN
matrices. It reverses each regular-Hadamard basis independently, then packs
the restored weights as MLX affine INT8/group-64.
transformer.conversion.json, SOURCE_MANIFEST.json, MODIFICATIONS.md, and
SHA256SUMS preserve the source, toolchain, transformation, and output
evidence. The conditioner, VAEs, and tokenizer are exact files from
Sawfwair/MiniMax-H3-FL2VA-MLX-4bit@e1244ad93d60c737c7e0f065a1c9372f3de7caf8.
Validation
The corrected artifact passed:
source-operation parity before MLX requantization (relative L2 ~= 1.7e-6);
packed transformer linear parity after MLX INT8 requantization
(relative L2 ~= 0.00733);
exact installed-cache parity against live three-modality AdaLN evaluation at
the released schedule point used by the regression gate (zero maximum output
error for both video and audio);
a coherent 256x160/22-frame native Ref2VA smoke;
a 512x320/124-frame native synchronized-A/V run on Apple M4 Max, producing
the included railway-platform example and an intelligible generated dialog
transcript.
The full validation run took 1,724.17 seconds with maximum acceleration. This
establishes correctness and usable output, not speed parity with FL2VA or LTX.
License
The model weights are governed by the MiniMax H3 Community License Agreement,
not the Apache-2.0 license of the mere.run source code. Preserve LICENSE,
NOTICE, and MODIFICATIONS.md, and comply with all territorial, downstream,
commercial, and acceptable-use requirements.
