hoidhxd/MiniMax-H3-x-Z-Image-bf16

MiniMax-H3 x Z-Image — BF16 ZS05

This repository contains a reconstructed BF16 FL2VA ZS05 diffusion model for the MiniMax-H3 x Z-Image workflow.

The model is reconstructed from the original pruned BF16 FL2VA model and the official/community ZS05 INT8 weights.

Files

FL2VA ZS05 BF16

minimax_h3_fl2va_pruned_zs05_bf16.safetensors

This is a BF16 reconstruction of the FL2VA ZS05 model.

The 21 ZS05 q_norm tensors in blocks 29–49 were copied directly from the ZS05 INT8 model after verifying that the INT8 and FP8 ZS05 versions produce the same BF16 q_norm reconstruction.

All non-target tensor payloads remain byte-for-byte identical to the original pruned FL2VA BF16 model.

REF2VA BF16

The corresponding REF2VA BF16 model is already available in:

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

Please download the REF2VA BF16 model from that repository rather than duplicating it here.

Why is this repository needed?

The original pruned FL2VA BF16 model does not contain the ZS05 q_norm values for blocks 29–49.

The ZS05 INT8/FP8 models contain the required ZS05 q_norm information.

Analysis showed that the INT8 and FP8 ZS05 versions have identical q_norm payloads after reconstruction, allowing the ZS05 q_norm tensors to be recovered exactly as BF16 values.

Only the 21 affected q_norm tensors were replaced.

No GPU conversion or approximate scalar/affine reconstruction is used in the final file.

Verification

Final verification performed:

• Tensor count: 532
• ZS05 q_norm tensors patched: 21
• Target tensors verified bit-exact: 21/21
• Non-target bytes unchanged: PASS
• Full verification: PASS

The resulting file is:

minimax_h3_fl2va_pruned_zs05_bf16.safetensors

Size:

~37.46 GiB

Hybrid B25-49

The BF16 FL2VA and REF2VA ZS05 models can be combined to create a hybrid B25-49 model.

The intended hybrid configuration is:

FL2VA ZS05 BF16
        +
REF2VA ZS05 BF16
        ↓
Hybrid B25-49 BF16

For the REF2VA component, use the BF16 model from:

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

Notes

This repository contains reconstructed weights intended for research and local inference workflows.

The ZS05 q_norm reconstruction changes only a very small amount of data (21 × 256 bytes = 5.25 KiB) compared with the ~37.46 GiB BF16 model.