I don’t recommend buying a discrete GPU. For home-level hardware iGPU systems like Strix Halo are much better than discrete systems. For example Strix Halo can run models that you need multiple 5090s for.

Not on top of my head, but there must be something. llama.cpp and vllm have basically solved the inference problem for LLMs. What you need is a RAG solution on top that also combines it with web search.

for coding tasks you need web search and RAG. It’s not the size of the model that matters, since even the largest models find solutions online.