Qwen 3.6 27B flags and performance in llama.cpp on RTX 5090

User reports Qwen 3.6 27B runs at 80-100 t/s on an RTX 5090 via llama.cpp, dropping to ~40 t/s at full 262k context. The setup barely fits in VRAM without vision and is used primarily for appdev tasks.