10-14-2026, 02:16 AM
Did a small comparison this afternoon. My result was mixed: first run looked bad, second run looked normal. The main thing that moved was latency, not quality.
2x3090 • llama.cpp • too much VRAM is still not enough