I benched Qwen 3.6 35B-A3B against Qwen 3.8 27B with the same parameters, thinking set to low. Despite 35B having 9x fewer active parameters, it benched only 2.34x slower. The 35B got only 50% more agentic tasks done per hour.