How Facebook-parent Meta’s AI engineers may be behind company delaying launch of Llama 4 ‘Behemoth’ model

Facebook parent Meta’s engineers may be responsible for the company postponing the launch of its Llama 4 ‘Behemoth’ AI model, a report claims. Sources familiar with the matter have suggested that the company’s engineers are reportedly facing challenges in improving the capabilities of the “Behemoth” flagship large-language model, ultimately delaying the rollout of the same. This has led the social media major’s staff to question whether the improvements of the upcoming model are substantial enough to need a public release. Early in its development, Behemoth was internally slated for an April release to coincide with Meta’s inaugural AI conference for developers. Before this event, Meta released two smaller models in its Llama AI family, but later pushed an internal target for the larger Behemoth’s release to June. Now, the release has been further delayed to fall or later, the Wall Street Journal reports.Meta CEO Mark Zuckerberg and other Meta executives haven’t publicly committed to a timeline for Behemoth, however, the company may decide to release a more limited version of the same sooner than expected. As per the report, Meta engineers and researchers may also be concerned about the AI model’s performance not matching public statements made about its capabilities.Meta's original Llama AI model was developed by its Fundamental AI Research Team, composed mostly of PhDs. Since the initial release in early 2023, 11 of the 14 original researchers have left. Subsequent versions, including Llama 4, were developed by a different team, whose performance has reportedly frustrated Meta's senior executives, the report adds.These issues have affected the progress of Meta's flagship AI model, Behemoth, despite public claims of its superiority over rivals. Internally, Behemoth has faced training difficulties, and Meta is considering major management changes, the report notes. Additionally, Meta also admitted to optimising its model specifically to perform well on a third-party benchmark test.