Skip to main content
Merge two compatible MiniMax H3 models by component, with a separate ratio for each part of the architecture. Each ratio is the share of model1, so you can keep the patch projections and refinement blocks of one model while replacing the transformer blocks with another.

Inputs

A ratio of 1.0 (the default) keeps that component entirely from model1 and 0.0 takes it entirely from model2; values in between blend the two. Tensors that carry no ratio, such as normalization statistics, always come from model1. Both inputs must be MiniMax H3 models with the same architecture, otherwise the merge produces a model that does not run.

Outputs

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 4fea4e588d16a8396c68aa071a59e671a2a0c645734606be6d7e221986ba88bd