[reddit r/LocalLLaMA] Long-context performance at lower quants