howtospark
Training

Model merging (mergekit)

Edit or adapt a modelMergingReference

Fuse multiple finetunes of the same base into one checkpoint by interpolating or sparsely merging their weights — no training, no extra params.

On the Spark

Merges are just tensor math on the 128 GB pool — the whole community's favorite way to make a new model without a GPU cluster.

Objective
Edit or adapt a model
Targets
Weights
Format
SLERP / TIES / DARE weight interpolation
Granularity
Whole-model, per-tensor weighting
Lifecycle
PTQ · no gradients
Calibration
No calibration
Compression
None — edits behavior, not size
Quality
Blends the parents' capabilities; quality depends on recipe
Hardware
None — offline weight edit, runs on any runtime
Runtimes

A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.