A custom hybrid version of MiniMax-H3 combining the FL2VA and REF2VA branches of the model.

The model uses different branches for different transformer blocks:

This hybrid layout was constructed at the raw safetensor payload level. No tensor conversion was performed when constructing the BF16 hybrid.

Approximately 37.46 GiB.

This is the full-precision BF16 hybrid model.

The B25-B49 tensors were copied byte-for-byte from the REF2VA ZS05 BF16 model, while the remaining tensors come from the FL2VA ZS05 BF16 model.

The repository also contains an INT8 version of the same hybrid model.

The INT8 model follows the same conceptual hybrid layout:

The BF16 hybrid was verified at tensor level.

Results:

All B25-B49 payloads were verified to be byte-exact copies from the REF2VA source.

No tensor conversion was performed during hybrid construction.

The purpose of this model is to experiment with a mixed FL2VA/REF2VA architecture where only the middle transformer blocks use REF2VA.

The selected layout is:

This allows the model to be tested as a single hybrid model without requiring separate model files for the two branches.

For the original/native Z-Image implementation and the REF2VA BF16 model, see:

Please refer to the original repository for the native model structure, configuration and usage information.

A GGUF conversion of the BF16 hybrid is provided separately:

The GGUF version is intended for use with GGUF-compatible ComfyUI workflows and loaders.

This is an experimental custom hybrid model.

It is not an official MiniMax release.

The hybrid block selection was created for experimentation and testing.