After years of using ChatGPT and Claude, I’m finally starting to dabble in local LLMs. I’m not replacing cloud AI yet, but running Qwen2.5 or Llama 3.2 on my PC comes in handy when I don’t want to hit message caps or censorship walls. You might think that running an LLM locally requires significant compute, but even an old GPU can handle several smaller models. I can’t expect Claude-like intelligence from my local AI setup, but it’s surprisingly good for general queries, document a***ysis, and productivity tasks, once I have the right tweaks dialed in. I was on the lookout for a cheap card to act as my dedicated local AI GPU, and was curious how low I could go. Fortunately, even an 8GB VRAM GPU that’s several generations old has enough power to host 7B–8B models. I got a pre-owned RTX 3060 12GB, and it cost me much less than upgrading my primary GPU to one of the latest high-end models.