One RTX4090 with 24G gives me about 200k KV cache with Q4_K_M quant for typica coding tasks, via opencode. I use Q6_K with 60k cache if the work doesn't require scanning too many files. Q6_K is really good.