Recipe Registry
Community Recipes
Battle-tested merge recipes with config.yaml files, full lineage graphs, and benchmark scores — ready to download and run.
7 recipes
Llama 3 × Gemma 2 Multilingual — Deep SLERP
A cross-family SLERP merge between Llama-3-8B-Instruct and Gemma-2-9B-IT, specifically tuned for multilingual output quality. The Gemma lineage brings Google's multilingual training signal; Llama brings instruction robustness.
Solar DARE — Llama 3 × Mistral DARE-TIES
DARE-TIES merge using random delta dropping to regularize a three-model combination of Llama-3 and two Mistral fine-tunes. Reduces interference between specialist models, producing one of the cleanest multi-model blends for general instruction following.
Qwen-Phi Coder — Task Arithmetic Specialist
Task arithmetic recipe subtracting a general-purpose vector from Qwen2.5-Coder and adding a math fine-tune task vector, surgically steering the model toward combined code + math excellence. Demonstrates the vector arithmetic approach to model editing.
Dolphin 2.9 — Llama 3 × Mistral SLERP Blend
A balanced SLERP merge of Dolphin Llama-3-8B-Instruct and Mistral-7B-Instruct-v0.3, producing a strong general-purpose assistant with excellent instruction following and broad world knowledge. Popular community recipe from r/LocalLLaMA.
OpenHermes Fusion — Mistral × Qwen TIES Multi-Model
A TIES merge of OpenHermes-2.5-Mistral-7B and Qwen2.5-7B-Instruct, targeting improved multilingual capability and coding while preserving Hermes instruction tuning. Sign conflicts resolved at density 0.7.
FrankenLlama 120B — Passthrough Layer Stack
A passthrough (frankenmerge) recipe stacking layers from Llama-3-70B and Llama-3-8B to produce a 120B architecture. Demonstrates the power of layer-level composition for creating models larger than any single input. Layers 0-39 from 70B, layers 40-79 spliced from 8B.
NeuralChat × DeepSeek — SLERP Instruct Blend
SLERP blend of Intel's NeuralChat-7B-v3.3 and DeepSeek-7B-Instruct at t=0.55, combining NeuralChat's strong instruction adherence with DeepSeek's reasoning depth. Well-regarded in the community for consistent benchmark improvements over either parent.