Comment by thot_experiment

Gemma 4 E4B is an incredible model for doing all the home assistant stuff I normally just used Qwen3.5 35BA4B + Whisper while leaving me with wayy more empty vram for other bullshit. It works as a drop in replacement for all of my "turn the lights off" or "when's the next train" type queries and does a good job of tool use. This is the really the first time vramlets get a model that's reliably day to day useful locally.

I'm curious/worried about the audio capability, I'm still using Whisper as the audio support hasn't landed in llama.cpp, and I'm not excited enough to temporarily rewire my stuff to use vLLM or whatever their reference impl is. The vision capabilities of Gemma are notably (thus far, could be impl specific issues?) much much worse than Qwen (even the big moe and dense gemma are much worse), hopefully the audio is at least on par with medium whisper.