MiniMax-H3-FL2VA-MLX-4bit

MiniMax-H3 FL2VA MLX 4-bit for mere.run

This is a compact, native-MLX FL2VA inference package for mere.run. It generates synchronized 24 fps video and 32 kHz stereo audio from text, a first frame, or directed first/last frames on Apple Silicon.

License and territory notice

These model files are governed by the included MiniMax-H3 Community License Agreement, not the mere.run source-code license. The license excludes use, distribution, and display in the United States, European Union, United Kingdom, and Republic of Korea and imposes downstream notice, modification, safety, and distribution obligations. Read LICENSE, NOTICE, and MODIFICATIONS.md before download, use, or redistribution.

By downloading, accessing, reproducing, or using these files, you acknowledge and accept the MiniMax-H3 Community License Agreement and agree to comply with it. The agreement does not authorize use, distribution, or display in the United States, European Union, United Kingdom, or Republic of Korea.

Provenance

Every tensor input came directly from the public official checkpoint:

MiniMaxAI/MiniMax-H3@ec19cc6daf5d8add9417c18e86b6b58cc6c55027

No converted or quantized third-party weights were used. The release was built with scripts/model-conversion/convert_minimax_h3_official_mlx.py from the mere.run repository in a declared Canadian conversion location. The converter verifies all 144.04 GB of official source files, checks the exact upstream license bytes, and requires MLX's CUDA Q4/Q8 packing to match byte-level fixtures captured on Apple Silicon before conversion.

Audit files:

• SOURCE_MANIFEST.json: byte count and SHA-256 for every official source file;
• transformer.conversion.json: converter, software, hardware, location, quantizer self-test, transforms, output sizes, and output hashes;
• SHA256SUMS: SHA-256 for every distributed runtime and legal file.

Numeric transforms

• 52 fused transformer/refiner QKV matrices: reordered from MiniMax's released per-head [q,k,v] row interleave into the official reference model's global [all-q; all-k; all-v] layout required by the native runtime;
• 208 active transformer/refiner core linears: official BF16 directly to MLX affine Q4, group size 64 after the QKV layout transform;
• precision-sensitive transformer projections, normalization, and outputs: retained at released BF16/F32 precision;
• exact 50-layer Qwen3-VL conditioner used by H3: 439 eligible linears from official BF16 directly to MLX affine Q8, group size 64;
• AdaLN cache: evaluated from the original official BF16/F32 projections over mere.run's released 31-point schedule before the cache-covered 13B-parameter inference branch was omitted;
• video VAE: official FP32 tensors cast directly to FP16;
• audio VAE: official FP32 weight-normalization pairs folded algebraically into equivalent plain convolution weights.

No weights were retrained, distilled, merged, or used as input from another quantized repository.

Use with mere.run

mere.run model pull video-minimax-h3-fl2va-mlx --accept-model-license
mere.run video generate \
  "a jeweled hummingbird hovering beside a red orchid, cinematic natural light" \
  --model video-minimax-h3-fl2va-mlx \
  --width 512 \
  --height 512 \
  --num-frames 56 \
  --seed 314159 \
  --output hummingbird.mp4

The managed package already contains its source-bound AdaLN cache; no post-download model optimize pass is required. Inference runs locally through native Swift/MLX. Python and CUDA are release-conversion tools only.

Validation

The complete 14-file managed bundle passed the fail-closed validator at 46,250,104,566 bytes. Anonymous Hub verification then matched the exact remote file set, public/ungated state, every LFS SHA-256 and size, every small-file hash, and the first and last 1 MiB of each large tensor.

A full-schedule generation was run locally on an Apple M4 Max with 128 GB of unified memory using the optimized arm64 release executable (SHA-256 042c25e023dddb71bb9fa44db0cca505dc4fe2115bcb17fd86a81953cebe4614):

mere.run video generate \
  "a jeweled hummingbird hovering beside a red orchid, cinematic natural light" \
  --model-root ./MiniMax-H3-FL2VA-MLX-4bit \
  --width 512 \
  --height 512 \
  --num-frames 56 \
  --fps 24 \
  --seed 314159 \
  --steps 31 \
  --h3-weight-mode quantized \
  --output hummingbird.mp4

• end-to-end elapsed time: 1,134.15 seconds (18:54.15);
• peak process footprint: 27,771,572,360 bytes, with zero swap and no thermal or performance warning;
• media: 512×512 H.264, 56 frames at 24 fps, plus 32 kHz stereo AAC, 2.334 seconds total;
• audio level: -26.1 dB mean / -7.8 dB peak;
• sample SHA-256: 08259a8c1a37a4947f3d8a3b329c63df59f6056a16698f595a817aeb2aaa0c6f;
• contact-sheet SHA-256: b6ac18a30fa35a1bfd67e9c8d265996a03e0447571ac32f369a83e29f76158dd.

Play the validated MP4 or inspect the six-frame contact sheet.