NEXS-multislerp-merges

NEXS Multi-SLERP Merges

A collection of four domain-expert model merges, one per architecture family, each produced by multi-SLERP in task-vector space with mergekit. Every merge is strictly intra-family — variants are only combined with other variants that share the same architecture and base model.

Method

Multi-SLERP (multislerp) performs barycentric spherical interpolation on a hypersphere for more than two models: it projects the models into the tangent space at their weighted Euclidean mean, interpolates, and projects back. Here it is run in task-vector space — each source's delta from the shared base model is computed, the deltas are spherically averaged with equal weight (normalize_weights: true, eps: 1e-8), and the result is added back to the base. Merging was done with mergekit.

For families whose variants had small vocabulary differences (a few added/removed special tokens), tokenizer_source: base aligns every variant's embed_tokens/lm_head to the shared base tokenizer before the merge, so all task vectors live in one embedding space.

Merged models