I tested Google's new Gemma 4 12B on my 8GB GPU, and now I don't want to go back to smaller models

A hands-on test of Google's Gemma 4 12B model running locally on a consumer 8GB GPU. The model features a novel architecture that skips separate encoders for multimodal input, a 256K context window, and native audio support. Setup involved trial and error in LM Studio with quantization settings to fit within GPU memory limits. Performance on reasoning, JSON generation, and document tasks was solid, outperforming smaller Gemma models for the author's workflow despite hardware constraints.