Can you merge AI models from different architecture families without retraining?
Models from different architecture families normally cannot be merged - tensor shapes and functional roles do not line up. VIDRAFT Chimera scores every tensor pair for compatibility, then separates what can be crossbred from what can only be transplanted.
Averaging weights across models of the same family is common. Across different families it does not work: layer counts, dimensions and the role each tensor plays all differ. Chimera scores every tensor pair for compatibility, sorts them into what can be crossbred and what can only be transplanted, and combines accordingly. Instead of blending parents into a diluted average, each parent core is frozen and only a small router is trained to pick the best parent per question - a 2-expert Chimera (4B) held GSM8K at 72.5%, identical to the parent, with 0.0 degradation. The method is covered by a filed patent with 29 claims.
Why can models from different families not simply be merged?
Within one family the architectures are close enough that averaging weights partly works. Across families the layer counts, dimensions and the role of each tensor all differ. Forcing an alignment makes the strengths that defined each model cancel one another out. That is why retraining is usually assumed to be necessary.
How does Chimera solve it?
Pairs are scored on shape, functional role and parameter distribution, then sorted: some can be combined, others must be moved intact. That is what turns the result from a lucky accident into a procedure. Structurally the attention is designed in-house while another model's FFN - the knowledge - is transplanted whole, and Mixture-of-Experts configurations are supported. The internal attention design and gating are not disclosed.
Does merging cost accuracy?
Blending parents like paint dilutes each strength. Instead every parent core is frozen and only a small router learns which parent to use per question. A 2-expert configuration (4B) held GSM8K at 72.5%, identical to the parent. Note that merging within one family gives a roughly neutral gain; the real lift comes from crossbreeding across families.
How much of this is public?
Compatibility-scored heterogeneous crossbreeding was filed on 8 May 2026 with 29 claims, as part of a 12-application portfolio totalling 317 claims. Measured results are published while the internal attention design, gating and merge recipe remain proprietary. Published figures are measurements under stated models and conditions and can change when conditions change.
Frequently asked questions
- Is model merging the same as model crossbreeding?
- They overlap but differ in scope. Common merging blends weights of same-family models. Crossbreeding extends to different families and includes the step of deciding what to combine versus what to transplant.
- Is retraining really unnecessary?
- The parent models themselves are not retrained. A small router that selects between parents is trained. That is a different order of cost from full retraining.
- What does GSM8K 72.5% mean here?
- It is the math accuracy measured on a 2-expert (4B) configuration, identical to the parent model. It means nothing was lost in merging - not that the result beats the parent.
Sources
Related
This article is based on VIDRAFT public and measured data. Performance figures are measurements under the stated models and conditions and can vary by environment. Internal methodology remains proprietary and patent-filed.