Looking for Suggestions — Single 5090 & 64gb DDR5

I say rule of thumb because since agentic flows/coding seem to surround like 90% of the conversations and these flows are often long context, multiturn, etc.and this is where KV cache quantization really shows its weaknesses. I've read enough anecdotal evidence on here and seen benches and papers showing fine quality degradation at longer sequences due to KV cache to know it's the last lever I touch when tuning. You're right though and apologies if my comment seemed condescending , it's not absolute and I appreciate the nuance that no config is one-size-fits all and everything has its tradeoff. As you state you need to find the ideal balance between both levers in your desired workflow within your VRAM and context budget. I would say Q4 + bf16 makes perfect sense in many flows, certainly more than the inverse. I wouldn't run Q4 for coding though (except perhaps a 200B+ model) but it works well in OCR/text based flows with lower parameter models. I certainly wouldn't advocate running a anything below a higher parameter Q3 model but that's just me, I personally don't bother but recognize my VRAM budget changes my perspective.