Gemma is MOE model, so not all parameters are activating at once. Qwen 27B is very demanding as all 27 billion parameters are active for every turn. This allows it to perform considerably better than anything in it's size as 99% of models are MOE these days.

CPU inference is going to be slow and you are going to be limited to very small MOE models. Give LM Studio a try, it is considerably better than Ollama.