KVarN, KV cache precision tail, low-bit quants in llama.cpp for longer context of better precision in the same VRAM