Minimax-H3-Hybrid-BF16-Pruned

MiniMax H3 Hybrid FL2VA + Ref2VA (B25–49, BF16 Pruned)

This repository contains a single-file, unquantized MiniMax H3 hybrid diffusion-transformer checkpoint for ComfyUI. It combines the higher-fidelity FL2VA checkpoint with the reference-conditioning pathway from Ref2VA.

The merge uses FL2VA as the base and replaces the per-block adaln_proj tensors in transformer blocks 25 through 49 with their Ref2VA counterparts. All other tensors remain from FL2VA. The result is intended to retain output quality closer to FL2VA while supporting image, video, and audio reference conditioning.

Checkpoint

This checkpoint is not INT8, FP8, GGUF, or otherwise quantized. The core model tensors are stored in BF16. A small number of auxiliary tensors retain the native FP16/FP32 dtypes of the source checkpoint.

Recommended use

Treat this hybrid as a Ref2VA checkpoint in loaders and inference runtimes. Although FL2VA supplies the base weights, Ref2VA request packing is required for multimodal reference inputs.

It is intended for:

• Reference-to-audio-video generation using image, video, and/or audio references.
• First/last-frame and text-driven generation where FL2VA-like fidelity is desirable.

Supporting MiniMax H3 components—text encoder, tokenizer/processor, visual VAE, and audio VAE—are not included in this repository.

Download

hf download dreamkrate/Minimax-H3-Hybrid-BF16-Pruned \
  minimax_h3_hybrid_fl2va_ref2va_b25-49-bf16-pruned.safetensors \
  --local-dir ComfyUI/models/diffusion_models

Then select the checkpoint in the MiniMax H3 diffusion-model loader.

How the merge was produced

The two official MiniMax H3 checkpoints share the same architecture and tensor layout. This checkpoint was made through deterministic tensor selection:

1. Start with the official FL2VA checkpoint as the base.
2. Copy adaln_proj weights from the official Ref2VA checkpoint for transformer blocks 25–49, inclusive.
3. Preserve all other tensors from FL2VA.
4. Export a self-contained, pruned safetensors checkpoint without quantization or additional training.

No fine-tuning, gradient updates, distillation, or interpolation were performed during this merge.

Tradeoffs and limitations

• This is an experimental merge. The block boundary was selected empirically and is not guaranteed to be optimal for every prompt or reference modality.
• Reference adherence may differ from the official Ref2VA checkpoint, while non-reference quality may differ from the official FL2VA checkpoint.
• No formal benchmark results are claimed. Evaluate the checkpoint on your own workflows and generation settings.
• This repository contains only the diffusion-transformer checkpoint, not the complete MiniMax H3 pipeline.

Provenance and acknowledgements

• Official base model: MiniMaxAI/MiniMax-H3
• Merge approach and comparison reference: smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models

All credit for MiniMax H3's architecture and original training belongs to MiniMax. This derivative checkpoint does not grant rights beyond the original MiniMax H3 license.

License

Use is subject to the MiniMax H3 Community License Agreement. Review that license before downloading, redistributing, or using the checkpoint.