Document measured MLX performance and tradeoffs

9cb13f7 verified 2 days ago

1.81 kB

# MiniMax H3 Image MLX · shared components

Shared VAE, PDD heads, native support tensors, transformer configuration and
evaluated six-step modulation tables for the community MLX port by Devin Lai.
Powered by MiniMax-H3-Image.

The [main runtime](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx) downloads these components once and reuses them
across BF16, lossless packed BF16, Q8 and Q4. It omits the redundant native
support file for the original BF16 variant. No other variant weights are
stored here. The separate native text encoder is fetched from its pinned
official `MiniMaxAI/MiniMax-H3` repository and retains its own license.

Use `h3-mlx download --variant packed12 --model-dir models/packed12` after
installation. Frozen modulation tables retain the six original timesteps and
the editing anchor; the loader validates MLX 0.32.4, strict FP32 settings and
the original schedule. [Format and modifications](MODIFICATIONS.md).

**Shared model files: 3.86 GiB.** Download them once and reuse them across variants. The BF16 downloader skips the redundant native-support copy; the separate upstream text encoder is roughly 62 GiB. These are download sizes, not RAM requirements.

This is an inference distribution, not a newly trained model. Training
provenance, biases and limitations remain those of the upstream model.
The [MiniMax-H3-Image Community License Agreement](LICENSE), territorial,
commercial and acceptable-use conditions apply. The standard grant excludes
the EU, UK, Republic of Korea and US. Manual access review does not grant
additional rights. [NOTICE](NOTICE).

**Release:** v0.4.1 (v0.4.0 weights with download-count metadata). Access requests are reviewed manually. [Main release status](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/RELEASE.md).

	# MiniMax H3 Image MLX · shared components

	Shared VAE, PDD heads, native support tensors, transformer configuration and
	evaluated six-step modulation tables for the community MLX port by Devin Lai.
	Powered by MiniMax-H3-Image.

	The [main runtime](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx) downloads these components once and reuses them
	across BF16, lossless packed BF16, Q8 and Q4. It omits the redundant native
	support file for the original BF16 variant. No other variant weights are
	stored here. The separate native text encoder is fetched from its pinned
	official `MiniMaxAI/MiniMax-H3` repository and retains its own license.

	Use `h3-mlx download --variant packed12 --model-dir models/packed12` after
	installation. Frozen modulation tables retain the six original timesteps and
	the editing anchor; the loader validates MLX 0.32.4, strict FP32 settings and
	the original schedule. [Format and modifications](MODIFICATIONS.md).

	Shared model files: 3.86 GiB. Download them once and reuse them across variants. The BF16 downloader skips the redundant native-support copy; the separate upstream text encoder is roughly 62 GiB. These are download sizes, not RAM requirements.

	This is an inference distribution, not a newly trained model. Training
	provenance, biases and limitations remain those of the upstream model.
	The [MiniMax-H3-Image Community License Agreement](LICENSE), territorial,
	commercial and acceptable-use conditions apply. The standard grant excludes
	the EU, UK, Republic of Korea and US. Manual access review does not grant
	additional rights. [NOTICE](NOTICE).

	Release: v0.4.1 (v0.4.0 weights with download-count metadata). Access requests are reviewed manually. [Main release status](https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/blob/main/RELEASE.md).