Qwen3.6-27B speculative decoding improves more on heavier quantizations

Benchmark on Qwen3.6-27B shows speculative decoding speed multiplier increases with heavier quantizations: Q8 > Q6 > Q4. Acceptance rate is quant-independent at matched depth, but base step slower on heavier quants.