Qwen3.6 users report reasoning loops

This is the messy underside of open-weight inference: better quant metrics do not automatically mean better wall-clock behavior, especially with reasoning mode, MoE routing, huge context, CPU offload, and fork-specific llama.cpp behavior in the mix.

• –Qwen3.6-35B-A3B is a serious open MoE model, but local serving stability still depends heavily on sampler settings, template handling, backend version, and quant choice
• –The user's config keeps reasoning on with unlimited budget, making repeated internal reasoning especially expensive when the model starts cycling
• –Q4_K_XL may preserve quality better than smaller IQ quants, but the speed drop from 40 tok/s to 27 tok/s can erase that benefit for interactive use
• –Recent community chatter around Qwen3.6 points to backend quirks in speculative decoding, tool calls, and recurrent-state handling, so upgrading llama.cpp/TurboQuant builds may matter as much as sampler tweaks