Access under the retained MiniMax agreements
Manual review, consistent with the other community weight releases. Read both MiniMax license files and their territory, commercial and acceptable-use conditions. Access approval grants no additional rights.
Log in or Sign Up to review the conditions and access this model content.
MiniMax H3 Image · ARA1 GGUF
Q8_0 diffusion transformer and conditioning encoder, with the original VAE, tokenizers, processor and six PDD heads. Powered by MiniMax-H3-Image.
Independent community conversion by Devin Lai.
This experimental release includes the published nine o_proj ARA edits in
encoder layers 31–39 at strength 1.0. No new training was performed.
Download
Request access on this page, read the retained model agreements and sign in. The complete weight bundle is 66.33 GB (61.77 GiB) before the included runtime source and documentation. Disk size is not the host RAM or GPU memory requirement.
python -m pip install 'huggingface-hub>=0.36.2'
hf auth login
hf download devin-lai/MiniMax-H3-Image-ARA1-GGUF \
--revision v1.0.0 --local-dir MiniMax-H3-Image-ARA1-GGUF
cd MiniMax-H3-Image-ARA1-GGUF
shasum -a 256 -c SHA256SUMS
| Component | File | Bytes |
|---|---|---|
| DiT | transformer/MiniMax-H3-Image-DiT-Q8_0.gguf |
36,022,938,432 |
| Encoder | text_encoder/MiniMax-H3-TextEncoder-ARA1-Q8_0.gguf |
27,921,295,104 |
| Original VAE | vae/diffusion_pytorch_model.safetensors |
2,336,964,672 |
The encoder contains the 50 language layers consumed by H3 conditioning and
the complete source-precision vision tower. The DiT uses the audited native QKV
row layout. Both component manifests record every tensor's shape, type and hash;
bundle.json records the packaged files. The GGUF weight bytes match the
previously validated local bundle exactly.
Run with the included Linux / NVIDIA runtime
runtime-source.tar.gz contains the patched SGLang backend, SageAttention
source, pinned dependencies and GGUF inference launchers. The GGUF loader fixes
are included; a generic SGLang installation may not contain them. This snapshot
contains inference sources, without the custom model export tools.
On Linux x86_64 with an allocated NVIDIA GPU, a C++ compiler, sufficient host RAM and a CUDA 13.0 compatible driver:
# Run from the downloaded model directory.
tar -xzf runtime-source.tar.gz
bash runtime/bin/install.sh
runtime/.venv/bin/python runtime/bin/serve_h3_gguf.py \
--bundle "$PWD" --gpu 0 --port 33000
In a second terminal, from the same model directory:
runtime/.venv/bin/python runtime/bin/generate.py \
'A watercolor fox under a snowy pine tree' \
--url http://127.0.0.1:33000 --profile six --quality exact \
--size 2048x2048 --seed 731 --output generated/fox.png
The service binds localhost. Keep ports 33000, 33001 and 33010 available. Weights stream from host RAM, with 12 DiT layers resident by default. Installation also needs scratch space for the pinned CUDA toolchain and compiled dependencies. The packaged sources correspond to the validated runtime; a new installation has not been rerun on a GPU as part of this upload.
Compatibility with ComfyUI and llama.cpp is untested. The GGUFs use H3-specific
tensor names and metadata; the .gguf extension alone does not establish loader
compatibility. The public MLX conversion scripts
document a separate storage-preserving repacking workflow for Apple silicon.
Validation and limits
Historical L20 tests used 24 prompts × two seeds at native 2K with all six PDD passes. Against the original BF16 image pipeline, mean LPIPS was 0.1798 and mean CLIP image cosine was 0.9784. The general-quality subset measured LPIPS 0.1448 and cosine 0.9848. These metrics do not establish perceptual or pixel equivalence; compositions can change substantially.
All three arms (original BF16, original Q8 control and edited Q8) generated 48/48 images with zero visually observed refusals. Refusal reduction was not demonstrated. The source language-model refusal figures are not image-generation results. Visual review was by Codex, without an independent human panel; the custom suite does not cover explicit sexual imagery or arbitrary prompts.
Standalone loading, 2K editing and native 4K generation were tested on L20. The recorded 4K check used 20,790 MiB peak PyTorch reserved GPU memory, excluding driver overhead. The final GGUF variant has not been validated on RTX 5090. See VALIDATION.txt and validation/ for the recorded methods, prompts, metrics and runtime checks.
Provenance and licenses
- Image source: MiniMaxAI/MiniMax-H3-Image, revision
6534ea1c979a5f8cb00c2bd1025d5ca4ee8d8901. - Conditioning source: MiniMaxAI/MiniMax-H3, revision
0919fb684102ef115126612f3fbb84d1027e7317. - Published edit source: llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic, revision
c44b949b30d111666a5ed9851c5cd633ed39b070.
Image weights retain the MiniMax-H3-Image Community License. The MiniMax-sourced conditioning encoder retains the MiniMax H3 Community License. Apache-2.0 attributions for Qwen/ARA materials and runtime code are separately retained in LICENSE-APACHE-2.0 and the runtime archive. Apache ancestry does not replace the MiniMax checkpoint license.
The model agreements include territory, commercial and acceptable-use conditions. These weights are released under those terms; they are not OSI open-source licensed. Access review follows the existing community weight releases and grants no additional rights. Retain all licenses, NOTICE and manifests when redistributing permitted copies. This is an independent derivative, without official MiniMax endorsement.
Release packaging adds the encoder agreement, corrects the historical encoder
license attribution in NOTICE and README.txt, and includes the runtime
snapshot. Model tensors are unchanged. The original manifest is retained in
provenance/original-bundle.json; the release manifest hashes the actual files.
- Downloads last month
- 2
8-bit
Model tree for devin-lai/MiniMax-H3-Image-ARA1-GGUF
Base model
MiniMaxAI/MiniMax-H3