Definition
Plain language
Combining several trained AI models into one, either by blending their internal numbers or by orchestrating their behavior.
As stated in the literature
Techniques for fusing multiple models; weight-level merging averages or stitches parameters and requires open internals, while behavioral-level merging (e.g., Fugu) combines closed models through their input-output behavior via routing and synthesis without touching weights.
Also called: weight merging, behavioral-level merging
Why it matters: It offers a way to combine capabilities without training a new model from scratch, including across models whose internal weights you can't access.
For example, you might blend a model good at coding with one good at writing to get a single model decent at both, or route questions between two closed models and stitch their replies together.
Heard on the show
“The world before this paper: multi-model coordination meant text passing or weight merging.”Episode 040 — Two Frozen Models Learn to Whisper: Coupling Through Hidden States