Llama 3.3 70B quantization hits multi-hop bottleneck

Llama 3.3 70B's high information density makes it uniquely fragile to aggressive quantization compared to its predecessors. Q4 variants consistently fail to integrate three or more pieces of information from different parts of the context, despite finding the correct chunks. Standard benchmarks like MMLU are failing to capture this "integration gap," which can mislead developers about real-world performance. The failure mode suggests quantization disproportionately damages the attention heads responsible for cross-section coherence. For complex RAG or agentic workflows, Q8 is now the mandatory floor for maintaining logical thread integrity. These findings, conducted on llama.cpp with a 16k context, indicate that quantization is no longer "basically free" for dense 70B models.