Qwen3.6-27B exposes local serving bottlenecks

This is less a model question than a serving question: once long-context workloads overlap, KV cache becomes the scarce resource, not raw token throughput.

• –vLLM is the right class of engine for this use case because its paged KV cache and prefix caching are designed to avoid recomputing shared prefixes and to manage concurrent requests more gracefully.
• –The user’s 80k-token bot checks are likely to trigger preemption or recomputation when memory is tight, which matches the failure mode vLLM documents for insufficient KV cache space.
• –36GB of total VRAM may still be tight for five active long-context requests once weights, activations, and cache overhead are all counted, especially if the workload really hits 120k contexts.
• –A modded RTX 3080 20GB helps only if the deployment plan changes with it; on its own, it is not a guarantee that five long-context sessions will stay resident without tradeoffs.
• –The cleaner win may be workload isolation: one inference instance for the bot, another for interactive coding, or a split across machines, rather than trying to make one box do everything.