Intermediate · Generative AI

Model Merging

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. Combining the weights of several fine-tuned checkpoints into a single model with no extra training, often via averaging in parameter space.

Technical Definition

Combining the weights of several fine-tuned checkpoints into a single model with no extra training, often via averaging in parameter space.

How it works

Methods include simple weight averaging (Model Soup), SLERP (spherical interpolation), TIES-Merging (resolve sign conflicts), DARE (drop-and-rescale), and Task Arithmetic (add/subtract task vectors). When the parent models share an initialization, merges often match or beat the best individual fine-tune at zero compute. mergekit is the standard toolkit and the technique dominates the open-LLM leaderboard.

Related Concepts

• Fine-Tuning — Adapting a pre-trained model to a specific task by continuing training on a smaller, task-specific dataset.
• LoRA (Low-Rank Adaptation) — A parameter-efficient fine-tuning method injecting small trainable matrices into frozen pre-trained layers.
• PEFT (Parameter-Efficient Fine-Tuning) — A family of methods that adapt large pretrained models by training only a small fraction of parameters, drastically cutting compute and storage.