>>106327320
>>106327374
>I want to be able to run some "decent" models
For decent MoE models, VRAM is not that important, you'll need more RAM instead.
For conventional dense models, even 3090 isn't that great because decent ones probably will need multiple of those to fit into VRAM entirely, with only one you'll have to run with partial CPU offloading one way or another.
For absolute poorfag setups even a tiny 4B model running on CPU could be considered 'decent'.

It's a matter of trade offs in the end, how slow of a generation speed you can tolerate, how low quality of a quantization you're going to run, etc.
I'm sitting here with 6600xt myself (8g of vram), something like qwen3 30b3ab runs 'fine'.

But for really peak stuff, there's no budget options.