You found a local AI app, you are finally ready to go, and then it hits you. Where do the actual models come from? They do not ship inside the app. You have to go get them, and there is a front door built for exactly that.
It is called a model hub, and it works like an app store for open weights. The big public one is Hugging Face, a giant open library where thousands of models live out in the open. Tools like Ollama also keep their own built-in registry, where you just pull a model by name and it downloads itself.
Now open any model page and read it like a pro. First, the family and the size. The family is the lineage, names like Llama or Mistral, and the size is how many parameters it packs, maybe eight billion. Bigger tends to be smarter, but it is also heavier.
Scroll down and you will find a whole stack of files for that same model. Those are quants, different compression levels, usually GGUF files at four bit, five bit, or eight bit. Do not just grab the biggest one out of pride. Grab the quant that actually fits your VRAM, so it runs fast instead of spilling onto slow system memory and crawling.
One last thing almost everyone skips: the license. It tells you whether you can use this model for a real business or just to tinker at home, and you really want to know that before you build on top of it. Family, size, quant, license. That is the front door. Off the Cloud.