SLERP and linear merges of Qwen2.5, Phi-4 and Chocolatine checkpoints for reasoning and multi-turn chat.