MiniMax-H3-Image-8bit

An MLX checkpoint converted directly from the original floating-point MiniMaxAI/MiniMax-H3-Image weights. It supports text-to-image generation and native image editing with up to sixteen ordered reference images. Inference uses MLX on Apple silicon, without PyTorch or Diffusers.

Quantization

Eligible linear and embedding layers use 8-bit affine quantization with group size 64. Both the diffusion transformer and the Qwen3-VL conditioner are quantized. The conditioner retains its vision tower for editing.

The VAE, transformer modulation layers, and transformer input/time/output projections retain their released precision. Normalizations and convolutional layers also retain their released precision.

Component	Weight files	Size (GB, decimal)
Qwen3-VL conditioner	6	27.491
Diffusion transformer	10	47.478
Image VAE	1	2.337

The full checkpoint occupies approximately 77.3 GB (72.0 GiB), including its tokenizer, configuration files, and weight indexes. No external image-model components are required.

Usage

Requires an mlx-vlm version with MiniMax-H3-Image support (implementation PR). Until that support is included in your installed release, install the implementation branch:

python -m pip install -U "git+https://github.com/lucasnewman/mlx-vlm.git@minimax-h3-image"

Replace /path/to/MiniMax-H3-Image-8bit with the local checkpoint directory or this repository's Hugging Face ID after upload.

mlx_vlm.generate --output-modality image \
  --model /path/to/MiniMax-H3-Image-8bit \
  --prompt "A red ceramic teapot beside a white cup on a wooden table by a sunny window" \
  --size 4096x4096 --seed 42 --output teapot.png

from mlx_vlm.generate import generate_image, load_image_generation_model

model = load_image_generation_model("/path/to/MiniMax-H3-Image-8bit")
result = generate_image(
    model,
    "A red ceramic teapot beside a white cup by a sunny window",
    width=4096,
    height=4096,
    seed=42,
    output_path="teapot.png",
)

For editing:

mlx_vlm.generate --output-modality image --task edit \
  --model /path/to/MiniMax-H3-Image-8bit \
  --image teapot.png \
  --prompt "Make the teapot in <Picture 1> blue, preserving its shape, the cup, and the lighting" \
  --size 4096x4096 --seed 42 --output blue-teapot.png

Reference images are ordered as <Picture 1>, <Picture 2>, and so on. Sampling is fixed at six steps and guidance 1. Small output requests are generated on at least a 2048x2048-equivalent canvas before resizing. The conditioner is released before diffusion, and the transformer caches fixed-schedule AdaLN projections so their large weights can be released.

Prompt expansion

Short prompts are expanded by default with the separately downloaded MiniMaxAI/MiniMax-H3-Image-PE model. Its released NVFP4 weights are separate from this 8-bit image checkpoint. Prompt templates and request defaults are embedded in the MLX implementation.

For a detailed caption that is ready for the image model, disable expansion:

mlx_vlm.generate --output-modality image \
  --model /path/to/MiniMax-H3-Image-8bit \
  --prompt "A realistic still-life photograph of a glossy red ceramic teapot with a curved spout and lid, beside a plain white cup on a warm wooden table next to a sunny window. Soft morning light from the left, detailed glaze and wood grain, a beige wall in the background." \
  --gen-kwargs '{"upsample_prompt":false}' \
  --size 4096x4096 --seed 42 --output teapot.png

In Python, pass upsample_prompt=False when loading or generating. With expansion disabled, this checkpoint can run fully offline after downloading.

Validation

Validated with strict component loading and offline inference on an Apple M5 Max with 128 GiB unified memory, using MLX 0.32.4. Both checks used seed 42, six steps, guidance 1, and prompt expansion disabled. The edit used the generated 4K image as its reference.