For quantized models, the quantization error propagates across the context size. I feel like 128k is the sweet spot preserving enough information for the current task in hand, and avoiding quality degradation due to quant error accumulation. I argue a 128k with compaction beats raw 256k/ 1M context
