“The Qwen team observed that certain model failure modes present in their 235 billion and 822 billion parameter models disappeared when scaling up to a 1 trillion parameter model.”