According to insiders, Meta has decided to delay the release of its highly anticipated next-generation large language model, Llama4, by at least until May this year. Although Zuckerberg has repeatedly expressed high expectations for the model, the development team encountered technical challenges in fine-tuning the base model's performance and optimizing logical reasoning, leading to the postponement of the original release plan.

As a cornerstone of Meta's core AI strategy, the progress of Llama4's development directly affects whether the company can gain an advantage over giants like OpenAI and Google. This delay reflects the increasing difficulty of balancing computing efficiency and complex instruction following in top-tier large models. Internal sources say that the team is using this extra time for more in-depth security stress testing, aiming to present a more dominant performance in the open-source community.

Although the release schedule has slowed down, Meta remains committed to its open-source strategy. It is reported that Llama4 will come in multiple versions with different parameter sizes to meet diverse inference needs, ranging from mobile devices to enterprise servers. For developers and researchers, although they need to wait two more months, more rigorous performance refinement may lead to more stable productivity results.

To support the massive training requirements, Meta continues to expand its high-end computing clusters. Although the release cycle has been extended, Meta emphasizes that its overall progress in multimodal understanding and long-text processing capabilities remains in line with expectations. The final unveiling in May will be a turning point to determine whether Meta can return to the top tier of global AI competitions.