User reports Qwen3.6 27B performs worse in vLLM than llama.cpp

A user with dual RTX 5060Ti (32GB VRAM) reports that Qwen3.6 27B produces dumber responses when run in vLLM compared to llama.cpp. The Reddit post has sparked discussion with 62 comments about potential causes like quantization differences or batching effects.