Qwen3.6-27B GGUF loops under llama-server

Hot take: the model is probably not “broken”; the serving stack is likely letting it dig its own hole.

• –Qwen3.6’s thinking mode is sensitive to template handling, and `preserve_thinking: true` can keep prior reasoning traces in play long enough to amplify repetition.
• –`repeat-penalty 1.0` is effectively no repetition penalty, so there’s little resistance once the model starts cycling.
• –`presence-penalty 0.0` and an aggressive speculative setup (`ngram-mod`) may make the loop harder to escape once the token stream drifts.
• –A cleaner test is to disable thinking preservation, add a real repeat penalty, and compare against a simpler llama.cpp template before blaming the GGUF.
• –The broader lesson: Qwen3.6 behaves more like a reasoning model than a plain chat model, so “works on my config” is very template-sensitive.