Qwen 3.6 27B speculative decoding benchmark on RTX 3090

Community benchmark tests Qwen 3.6 27B with speculative decoding, achieving ~100 tokens per second on a single RTX 3090. Compares 5 engines including llama.cpp forks and Lucebox.