“Models from Character.ai and Google's Gemma series employ an architecture where the global context (KV cache) is shared across all layers.”