What if you could combine the strengths of two models without retraining? Create a model that writes code like Model A but reasons like Model B? This is model merging, and it works.<\/p>\n
Model merging takes weights from two or more models and combines them in clever ways. The result is often surprising: emergent capabilities you wouldn’t expect from simple averaging.<\/p>\n
The simplest merge is linear interpolation. If Model A has weights W_A and Model B has weights W_B, the merged model has weights W = (1-a)*W_A + a*W_B for some weight a.<\/p>\n
This almost never works well. Naive averaging destroys the delicate weight distributions both models learned. But with careful techniques, it works surprisingly well.<\/p>\n
SLERP (Spherical Linear Interpolation) treats weight vectors as points on a sphere. Instead of straight-line interpolation, it moves along a geodesic through the weight space.<\/p>\n
SLERP preserves the magnitude of weight vectors better than linear interpolation. The result: merges that maintain model coherence better.<\/p>\n
TIES (Trim, Interleave, and Ensemble) is more sophisticated. It identifies the most important weight changes in each model, combines only those changes, and uses ensemble techniques to blend results.<\/p>\n
TIES has published results showing that merging a code model with a reasoning model produces better performance on tasks requiring both skills than either model alone.<\/p>\n
DARE (Domain Adaptation and Rapid Ensemble) randomly samples weights from each model instead of averaging. Counter-intuitively, this works well for merging models fine-tuned on different datasets.<\/p>\n
DARE is particularly good for combining multiple fine-tuned models (e.g., 5 different LoRA adapters) into a single coherent model.<\/p>\n
The key insight is that fine-tuned models share the same base architecture and are trained from the same initialization. Their weight spaces are aligned in ways that allow meaningful interpolation.<\/p>\n
When you merge models that diverged from the same starting point, you’re not combining arbitrary weight matrices. You’re blending carefully learned deviations from a common base.<\/p>\n
Combining specialized adapters:<\/strong> Train 5 LoRA adapters on different domains, merge them into a single multi-domain model. Rapid model development:<\/strong> Don’t have time to train? Merge two existing models and iterate from there.<\/p>\n mergekit is the standard tool. It handles SLERP, TIES, DARE, and custom merge strategies. Using it is trivial:<\/p>\n Define a YAML config specifying which models to merge and which method. Run mergekit. Get a merged model.<\/p>\n The process is fast (minutes, not hours) and requires no training.<\/p>\n Merging only works well when models are compatible: same architecture, similar capability levels, trained from the same initialization.<\/p>\n Merging a 7B and a 70B model won’t work. Merging models from different architectures won’t work. But within compatible families, merging is powerful.<\/p>\n Model merging democratizes the ability to create specialized models. You don’t need to train from scratch. Combine existing models, and you often get something better than any individual model.<\/p>\n This is particularly powerful in the era of open source models where dozens of fine-tuned variants exist for every task.<\/p>\n
\nBalancing trade-offs:<\/strong> One model is verbose but accurate. Another is concise but sometimes wrong. Merge them to balance both.<\/p>\nTools for Merging<\/h2>\n
The Limitation<\/h2>\n
What This Means<\/h2>\n