Skip to content

v0.5.0 · MiniMax H3 Image on Apple silicon

Latest

Choose a tag to compare

@devin-lai devin-lai released this 13 Nov 02:32

Generate and edit images locally on Apple silicon with MLX and Metal. This is the GitHub inference release for the MiniMax H3 Image MLX collection.

Inference-only, Apache-2.0 code. Model conversion, packing encoders, online quantization and weight-export tools are not included. Model weights retain their separate MiniMax agreements.

Included

  • Six-step generation and reference-image editing.
  • Pinned BF16, lossless packed BF16, Q8 and Q4 downloads, shared-file reuse and optional SHA-256 verification.
  • Resident pipelines, JSON inference records, saved trajectories and exact M5 tuning with native fallbacks.
  • English/Chinese docs, recorded sample prompts/seeds/hashes, benchmark comparisons and an inference benchmark runner.

Measured performance

2048 × 2048, six steps, M5 Pro / 16 GPU cores / 64 GiB, MLX 0.32.4:

Storage Warm control Warm tuned Reduction
BF16 65.83 s 62.49 s 5.07%
Lossless packed BF16 66.42 s 64.05 s 3.57%

Historical paired campaigns with cached conditioning and resident weights. Geometric mean of two case medians, three alternating repeats per case. Startup/downloads excluded; trace and PNG save included. Separate campaigns, each compared with its own MLX control. These are not cold-start or CUDA speedup claims. Method and raw records.

Validation

61 local tests pass. Source, wheel and sdist pass the inference-only audit. Installed-wheel BF16, packed12 and Q8 generations reproduce the recorded PNG hashes and all 16 saved trajectory tensors. Q4 has loader tests, pinned manifest checks and separately labeled historical image evidence; it was not freshly regenerated. Validation receipts.

Get started

uv pip install 'minimax-h3-image-mlx[native] @ git+https://github.com/devin-lai/MiniMax-H3-Image-MLX.git@v0.5.0'
h3-mlx doctor
hf auth login
h3-mlx download --variant packed12 --model-dir models/packed12

Quick start · Usage · Recorded AI-generated samples

Request model access on Hugging Face before downloading. The sample gallery retains its existing access review under the model terms. Assets below contain no model weights or original sample PNGs; verify them with SHA256SUMS. Existing Hub model revisions remain unchanged.

Independent community port by Devin Lai. Powered by MiniMax-H3-Image.