I tested 3 local LLMs on my RTX 4070 Ti for real work — only one earned a permanent spot

12GB VRAM, one model worth keeping.