A Reproducible Within-Family SLERP Merge of Llama-3.1-8B-Instruct and DeepSeek-R1-Distill-Llama-8B

We document a reproducible weight-level merge of two same-family checkpoints (Llama-3.1-8B-Instruct and DeepSeek-R1-Distill-Llama-8B) using Spherical Linear Interpolation (SLERP) at t=0.5. The merge is performed via the open-source mergekit toolkit on a 188 GB CPU server in approximately three minutes. The resulting 8B-parameter model loads cleanly through the Hugging Face transformers library and generates coherent English on five out-of-distribution prompts spanning explanation, arithmetic, poetry, factual recall, and algebra. We then quantitatively benchmark the merged model against both base checkpoints on a 10-prompt held-out GSM8K subset. The merged artifact achieves 60% accuracy matching Llama-3.1-8B-Instruct baseline (zero merge tax in this configuration) while acquiring the reasoning-marker style of the DeepSeek-R1-Distill donor (4.30 markers per generation vs 2.45 for Llama). We document a caveat: DeepSeek-R1-Distill itself scored only 10% under our 200-token budget, reflecting incompatibility between the budget and R1-family reasoning chain length, not the model's underlying capability. We pair this result with a companion negative finding on activation-level function vector injection between the same two checkpoints: weight-level SLERP produces a usable composed model that inherits stylistic features from the donor without accuracy loss, while naive single-layer activation steering does not transfer the donor's reasoning capability. This contrast motivates a layered framework in which different composition mechanisms address different sub-problems.