devin-lai's picture
Document measured MLX performance and tradeoffs
9cb13f7 verified
|
Raw History Blame Contribute Delete
3.58 kB
# MiniMax H3 Image MLX 路 Q4 affine core
Fifty main DiT blocks and two token-refiner blocks use MLX affine 4-bit weight quantization, group size 64. This is the smallest core and has stronger observed lettering and detail drift. Inspect the comparisons before choosing it.
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) 路 [Compare variants](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/MODEL_INDEX.md) 路 [Evidence and limitations](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/BENCHMARKS.md)
## Storage and historical performance 路 M5 Pro / 64 GiB
| Measure | Result |
|---|---:|
| Q4 variant payload | **10.50 GiB** |
| One historical warm product trial: denoising | 67.71 s |
| Same trial: peak MLX allocation | 12.51 GiB |
The variant payload excludes approximately 3.86 GiB of shared components and the roughly 62 GiB text encoder. The historical trial was one 2048 脳 2048, six-step product run with cached conditioning and resident weights, using online Q4 on an M5 Pro / 64 GiB. Denoising is not end-to-end time; MLX allocation is not total process or system memory and does not establish minimum RAM. This was not a fresh performance test of the downloadable export. Q4 changes images, especially detail and lettering, with no established speed advantage. [Method and full comparisons](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/BENCHMARKS.md#historical-q8-and-q4-measurements).
## Download only this variant
After installing the main runtime and obtaining the required access:
```bash
h3-mlx download --variant q4 --model-dir models/q4
h3-mlx generate --model models/q4 \
--text-encoder models/q4/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) 路 [Follow devin-lai](https://huggingface.co/devin-lai)