[deviation]

Their tokenization suggests they're new Qwen models AFAIK. They tokenize input to the exact same # of tokens that Qwen models do.