Qwen3.6 hangs on dual 5060 Ti

This reads like a serving/configuration problem, not a PC failure. A 27B model can absolutely run locally, but asking it to carry a 100k-token context on consumer GPUs pushes the backend into the zone where it looks frozen even while the OS stays responsive.

• –Qwen3.6-27B is a dense, long-context model, so the model itself is not the problem; the memory footprint of the active context is.
• –`llama.cpp` and Ollama both have to manage KV cache and prompt reprocessing, so each follow-up can get more expensive than the last if the conversation history keeps growing.
• –OpenWebUI plus search/RAG can silently inflate prompts, which makes the third or fourth turn much heavier than the first.
• –Two GPUs do not automatically mean better stability or throughput; sharding can add overhead, especially if the model already fits poorly with the chosen context size.
• –The practical fix is to cut context size aggressively, trim history, and test a simpler single-backend setup before assuming a hardware fault.