Stack layers from different models into a single new network — no retraining, just surgery on the layer stack.
What it is
Where a model soup averages weights, frankenmerging splices architecture: take blocks of layers from two or more models and concatenate them into one taller network. The stitched model can pick up behaviors from each donor, sometimes producing capabilities none of the parents had alone.
It is the most literal form of model composition — building a new brain out of parts of other brains — and, being training-free, the cheapest to experiment with.
Why it's worth watching
Part of the fast-maturing model-merging field (mergekit popularized passthrough/frankenmerge; covered in the 2026 ACM Computing Surveys review of merging).
Obscure enough as a coinage that it remains uncontested — but the underlying practice is spreading through the open-weights community.
The composition angle
Frankenmerging is composition at the structural level. Time, weights, structure — three axes, one thesis: the winning systems are composed from small specialized pieces, not carved from one giant model.