Model merging combines weights from multiple fine-tuned models to create a single model with combined capabilities - no additional training or GPU time required. Why it works: fine-tuned models start from same base, their weight differences encode capability improvements, interpolating in weight space can preserve both improvements. Core techniques: 1) Simple averaging (Model Soup): average weights of multiple fine-tuned models. Surprisingly effective when models have similar base, improves generalization and reduces overfitting. 2) SLERP (Spherical Linear Interpolation): interpolate on sphere rather than linearly, preserves magnitude of weight vectors. Better for models fine-tuned for very different tasks. 3) TIES-Merging: addresses sign conflicts between models - first trim low-magnitude changes, then resolve sign conflicts by majority vote, then merge. 4) DARE (Drop and Rescue): randomly drop fine-tuning changes, scale remaining changes to preserve expected magnitude. Reduces interference between models. 5) Model Frankensteining: take different layers from different models (e.g., first 16 layers from model A, last 16 from model B). Tools: mergekit by Charles Goddard enables command-line model merging with all these techniques. Practical application: merge domain-specific fine-tune with instruction-following fine-tune to get model good at both.