MTP speed falls off past 85K context

This is a useful reality check for long-context local inference: the feature works, but the tail latency and throughput curve still bend sharply once the session gets really long.

• –Performance degradation past 85K context suggests the practical ceiling for “daily driver” coding sessions is lower than the raw context window implies
• –Cold prefill cost is still the main tax for new sessions, so reuse and cache persistence matter a lot more than marketing benchmarks
• –KV cache slot-save looks like the unsung hero here; improving hit rate is probably more valuable than chasing small decode gains
• –Qwen3.6-27B Q4_K_M remains viable for local coding, but this session shows why observability matters more than vibes when you push long contexts
• –The post is more of an engineering benchmark note than a launch: it helps separate “usable in practice” from “works on paper”