Meta has denied allegations that it manipulated its latest AI models, Llama 4 Maverick and Llama 4 Scout, to achieve better benchmark scores while hiding their limitations.

Ahmad Al-Dahle, Meta’s vice president of generative AI, addressed the rumors in a post on X, saying they are “simply not true.” He rejected claims that the company trained the models on “test sets,” which are typically used for performance evaluation after training. This could result in misleading scores.