Aratako/LightChatAssistant-2x7B-optimized-experimental
MoE built from dense models vs Elizezen/Antler-7B. Aratako/LightChatAssistant-2x7B-optimized-experimental is a mixture of 2 experts with the same shape as Elizezen/Antler-7B's MLP, but none of them is a copy of it: they come from other models or were trained further. Shared tensors (attention, norms, embeddings): 11 of 195 identical.