Multi-model serverless can be markedly slower than specialized fast hosts Connections part_of Large Language Model uses Gemma 4
