arcee-ai/mergekit

Supports multiple merge algorithms (linear interpolation, SLERP, task arithmetic, expert composition) and specialized techniques like LoRA extraction, layer-wise "Frankenmerging," and Mixture of Experts assembly. Uses an out-of-core lazy-loading approach to operate on CPU or minimal VRAM (8GB), enabling complex merges in resource-constrained environments. Integrates with Hugging Face Hub for model distribution and accepts YAML configurations for flexible multi-stage merge workflows across Llama, Mistral, GPT-NeoX, and other architectures.

6,857 stars. Actively maintained with 1 commit in the last 30 days.