I’m using Gemma 4 31B right now, and for my use cases I actually prefer it over Qwen 3.5 122B.

What stands out to me is that Gemma 4 31B tends to get what I’m asking for right on the first try, whereas Qwen will often get there eventually but usually needs more steering.

As for prompt processing, I haven’t measured it yet, but they seem fairly similar to me, maybe with a slight edge to Qwen. In the decoding phase, Qwen also feels faster. But honestly, Gemma’s outputs have been good enough that I still prefer it overall.

Gemma can be pretty aggressive with RAM usage. It may start off lighter, but once you actually work with it for a while, you can watch the RAM disappear. My understanding is that this is tied to how Gemma handles KV cache. I’ve heard turbo quant is strongly recommended to help reduce KV-cache memory usage, but I haven’t implemented it yet. Even so, I’m convinced enough by Gemma 4 that it’s an optimization path I’m willing to pursue.