MiniMax-H3-Pruned-full (self-contained mirror)

A self-contained re-packaging of the pruned MiniMax H3 checkpoint published by multimodalart, with the components that the pruned index borrows from the base release copied in, so that one repository holds everything the diffusers modular pipeline needs. This exists because some serving platforms mount exactly one Hugging Face repository per endpoint.

Component	Source	Pinned commit
text_encoder (Qwen3-VL, 51 layers), transformer, transformer_ref, remote code modeling_minimax_h3_pruned.py	multimodalart/MiniMax-H3-Pruned	1a0ef5e65b639e84af81d883817968532180e9c7
vae, audio_vae, tokenizer, processor, scheduler, audio_scheduler, LICENSE	MiniMaxAI/MiniMax-H3	9ac0dd7aabc2c651fcf0ace4c00b2bffd9c8c8a6

modular_model_index.json is the upstream pruned index with every pretrained_model_name_or_path rewritten to lightweight/MiniMax-H3-Pruned-full; no component resolves to another repository at load time. All weights are byte-identical to the sources. Loading requires trust_remote_code=True (pruned transformer class). Upstream pruned model card: README_upstream_pruned.md.

License

Use is subject to the MiniMax H3 Community License Agreement (LICENSE), including its Acceptable Use Policy and its territorial scope. Credit to MiniMax for the model and to multimodalart for the pruned checkpoint.