MiniMax-H3 Ref2VA MLX 8-bit validation

This receipt records the corrected managed Ref2VA artifact and one real native Swift/MLX generation on an Apple M4 Max MacBook Pro with 128 GB unified memory. It is correctness and usability evidence, not a claim of speed parity with FL2VA or LTX.

Artifact identity

• Managed ID: video-minimax-h3-ref2va-mlx
• Repository: Sawfwair/MiniMax-H3-Ref2VA-MLX-8bit
• Immutable revision: 61dc387ef1a7166425cdacd63c2340598dcc364f
• Managed bytes: 70,941,103,245
• Source: Comfy-Org/MiniMax-H3@fd70b39279d1ae6eb214c903f53e1bec3af19a77
• Source file bytes: 34,038,894,550
• Source SHA-256: 9eef934046a0671bc8a5daf87100705e1478419c574cfde70c50fbe6885f76a9
• Converted transformer bytes: 36,024,412,656
• Converted transformer SHA-256:234f22f69f8d40d6ed81cceed8259fa287f3c9417d40fba5274e3a7aa84e18a2

The transformer and Qwen3-VL conditioner are MLX affine INT8/group-64. The video VAE is FP16 and the audio VAE is FP32. Lower Ref2VA precision did not meet the visual quality bar, so 8-bit is the published floor.

The 14-file managed pull also includes adaln_cache.safetensors, exactly 873,820,740 bytes with SHA-256 2cbe9e3324ef2cc5108a3ba7f1219d84079ff00a017f604fd86300005cc64fcd. Its source identity is the immutable converted-transformer SHA-256. The cache was built one three-modality timestep batch per released schedule point, then reloaded and compared with both a fresh cache and direct live AdaLN evaluation. At schedule step 10, maximum video and audio output error were both zero.

Managed revision upgrade

The artifact was first installed at pre-cache revision abb9114fe9d6e3cccc6376eee1abaf09d3f2a9fe, with the corrected cache added to the model root locally. After the catalog pin moved to the recorded revision, model pull treats the earlier manifest as stale and asks the Hub tree for the target revision's Git or LFS object identities. It preserves the installed root while preparing the immutable snapshot and adopts a local payload only after its byte count and Git blob SHA-1 or LFS SHA-256 match the target object.

On that real stale install, structured preflight reported 4,362 bytes rather than the package's 70.94 GB logical size. Those bytes were the two changed managed provenance files; every large target object, including the 873,820,740-byte cache, was already available by exact content identity. A pull, without --force or a preparation command, produced a 14-file snapshot receipt whose requested and resolved revision are both 61dc387ef1a7166425cdacd63c2340598dcc364f. Post-pull model validation checks the cache byte count and safetensors format, schema, and transformer-bound source identity in addition to the conversion receipt.

Conversion correction

The source checkpoint stores per-tensor ConvRot metadata. Two different source rotation groups are present: group 256 for 200 transformer matrices and group 64 for 50 AdaLN matrices. The faulty conversion treated the MLX output group size as though it were also the source rotation group, producing noise.

The corrected converter reads and validates every tensor's embedded source group, reverses that regular-Hadamard basis, and only then requantizes the restored matrix to MLX affine INT8/group-64. Unit fixtures cover both source groups and mixed-group conversion. The final artifact records the toolchain, source identity, group counts, quantization, output size, and hash in transformer.conversion.json.

Numerical checks on the corrected route measured approximately 1.729e-6 relative L2 error before MLX requantization and 0.0073349 for a packed linear operation after INT8 requantization.

Native generation receipt

The full validation used real reference conditioning and maximum acceleration. The run produced these results:

• Video: 512 x 320 H.264.
• Duration: 124 frames at 24 frames per second (5.167 seconds).
• Audio: AAC stereo at 32 kHz.
• Wall time: 1,724.17 seconds.
• MP4 SHA-256:08ce4cfb9fe305ba67297d94f6a54f3ce49c12bb8e242c38561cb6cb6237c9b0.

The output was visually coherent, retained usable motion and subject structure, and generated intelligible synchronized dialogue transcribed as: “You kept up the recording? Yeah. Every second.” A preceding 256 x 160, 22-frame smoke was also coherent. This establishes that the corrected artifact executes the real Ref2VA path without the all-noise failure.

The 28-minute wall time confirms that this path remains substantially slower than desirable. Treat the result as a quality and conditioning candidate for an LTX and H3 comparison, and measure performance separately.