# MiniMax H3 Image MLX · Lossless packed BF16

Recommended exact storage choice. Fifty main DiT blocks use the lossless palette12 codec: original BF16 payloads are reconstructed before GEMM. This is not 12-bit quantization. Native token refiners and support tensors come from the shared repository.

Community adaptation by **Devin Lai**. Powered by MiniMax-H3-Image.

[Start here: install and quick start](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx) · [Compare variants](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/MODEL_INDEX.md) · [Evidence and limitations](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/BENCHMARKS.md)

## Measured performance · M5 Pro / 64 GiB

| Measure | Result |
|---|---:|
| Warm 2048 × 2048 image, tuned | **64.05 s** |
| Matched packed control → tuned | 66.42 s → 64.05 s (**3.57% lower**) |
| Maximum sampled parent footprint | **32.56 GiB** |
| Packed variant payload | **27.23 GiB** |

The payload excludes approximately 3.86 GiB of shared components and the roughly 62 GiB text encoder. Packing reconstructs the original BF16 payloads; the four recorded validation cases matched the corresponding MLX controls pixel-for-pixel and in saved denoising tensors.

Warm 2048 × 2048 generation with all six steps, cached conditioning and resident weights. The result is the geometric mean of product and portrait medians, each with three alternating repeats. Startup and downloads are excluded. Peak is sampled parent-process physical footprint across four candidate cases, not a minimum memory requirement. The tested device was an M5 Pro with 16 GPU cores and 64 GiB, running MLX 0.32.4.

The optimized settings are optional and a new CLI process includes loading. [Tuning flags](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/USAGE.md#measured-exact-tuning) · [Method and receipts](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/BENCHMARKS.md#exact-tuning-results).

## Download only this variant

After installing the main runtime and obtaining the required access:

```bash
h3-mlx download --variant packed12 --model-dir models/packed12
h3-mlx generate --model models/packed12 \
  --text-encoder models/packed12/text_encoder \
  'An editorial photograph of a cobalt perfume bottle on pale stone' \
  --seed 731 --output outputs/bottle.png
```

This fetches this variant, required shared components and the separately pinned
native text encoder (roughly 62 GiB). It does not fetch other variants. Use
`--text-encoder-path /path/to/text_encoder` to reuse an existing encoder.
All image variants retain the six original PDD heads and native FP32 islands.
The published compact format requires this runtime, not a generic Diffusers
or mlx-lm loader. Tested on M5 Pro / 64 GiB; smaller Macs are unqualified.

## Model card and terms

This is an inference port of `MiniMaxAI/MiniMax-H3-Image` at revision
`6534ea1c979a5f8cb00c2bd1025d5ca4ee8d8901`, not a newly trained model.
No new training dataset is introduced. Training provenance, potential biases,
intended uses and upstream limitations remain those of the original model.
See [modifications](MODIFICATIONS.md), [license](LICENSE) and [NOTICE](NOTICE).

The MiniMax-H3-Image Community License Agreement and its acceptable-use policy
apply, including territorial and commercial restrictions. The standard grant
excludes the EU, UK, Republic of Korea and US. Access is subject to manual
review; it does not provide separate permission. The upstream text encoder
retains its own MiniMax H3 agreement. This is not an unrestricted open-source
license or an official MiniMax endorsement.

**Release:** v0.4.1 (v0.4.0 weights with download-count metadata). Access requests are reviewed manually under the retained license.
The actual file manifest records exact sizes and SHA-256 hashes.

[Report reproducible results or issues](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/discussions) · [Follow devin-lai](https://huggingface.co/devin-lai)