KV cache quantization degrades DeepSeek V4 Flash quality, test finds

A Reddit user reports that switching DeepSeek V4 Flash's KV cache from BF16 to Q8 significantly degrades quality across perplexity, KL divergence, and Same TopP metrics. The effect stands in contrast to Qwen 397B, where quantization impact is smaller.