Qwen 3.6 27B speculative decoding benchmark achieves ~100 TPS on RTX 3090

Benchmark shows ~100 tokens per second on a single RTX 3090. Tested 5 engines including llama.cpp forks and Lucebox with detailed throughput comparisons.