Gemma4 2B is the smallest of the Gemmas. Google made it with "efficient architecture", also, which I am told makes it more effective on less powerful devices, like phones. The difference is about 3gb of filesize, and from what I can tell from using it it is only marginally slower than Qwen3 1.7b (but way more capable). It still runs fine on Raspberry Pi 5.
It did not work correctly on a smaller LLM, but without deep insight into how the LLMs are constructed it's really hard to tell what they're fully capable of. It's actually crazy how different Qwen and Gemma are in inference styles. I really like Qwen, it's a very straightforward sort of LLM. Gemma4 is somewhat rambling. I'm certain that in future we could get this sort of thing down to a sub-1B model, just not right now.