Great project! How does this compare to asking qwen to reply with 1 token in terms of speed?