hoidhxd/MiniMax-H3-x-Z-Image-hybrid-GGUF

MiniMax-H3 x Z-Image Hybrid GGUF

GGUF conversion of the custom MiniMax-H3 x Z-Image Hybrid model.

The underlying hybrid architecture is:

B00-B24  -> FL2VA
B25-B49  -> REF2VA
B50+     -> FL2VA

This model combines the FL2VA and REF2VA branches into a single hybrid model.

Files

BF16 GGUF

minimax_h3_hybrid_fl2va_ref2va_zs05_b25-49_BF16.gguf

This file was converted from:

minimax_h3_hybrid_fl2va_ref2va_zs05_b25-49_bf16.safetensors

using the molbal/ComfyUI-GGUF conversion tools.

The BF16 GGUF is provided as the source GGUF representation before additional quantization.

Quantization

Additional quantized GGUF versions may be added later.

For example:

Q4_0
Q5_0
Q8_0

will be added as separate files when available and verified.

Original Hybrid Model

The original BF16/INT8 safetensor versions are available here:

https://huggingface.co/hoidhxd/MiniMax-H3-x-Z-Image-hybrid

Native / Reference Repository

For the original/native MiniMax-H3 x Z-Image model and REF2VA files, see:

https://huggingface.co/joeygambino/MiniMax-H3-x-Z-Image-native

ComfyUI

This GGUF model is intended for use with:

https://github.com/molbal/ComfyUI-GGUF

The conversion used the molbal/ComfyUI-GGUF implementation.

Make sure your ComfyUI-GGUF version supports the MiniMax-H3 architecture.

Hybrid Layout

The model is intentionally constructed as:

                    MiniMax-H3 Hybrid

        ┌───────────────────────────────────┐
        │ B00 - B24                         │
        │ FL2VA                             │
        ├───────────────────────────────────┤
        │ B25 - B49                         │
        │ REF2VA                            │
        ├───────────────────────────────────┤
        │ B50+                              │
        │ FL2VA                             │
        └───────────────────────────────────┘

Verification

The original BF16 safetensor hybrid was verified at tensor level:

Tensor count                 : 532
Expected source matches      : 532/532
Payload failures             : 0
Wrong-source tensors         : 0

B00-B24 : FL2VA
B25-B49 : REF2VA
B50+    : FL2VA

Tensor set       : PASS
dtype            : PASS
shape            : PASS
payload sizes    : PASS
Payload hashes   : PASS

The hybrid construction itself was performed by byte-level tensor replacement. No BF16 numerical conversion was performed.

Important

This is an experimental community-created hybrid.

It is not an official MiniMax release.

Use the native repository for the original model and official model structure.