Why don't they adapt the deepseek v4.1flash model instead?