Meta is preparing to release the latest version of its large language model, Llama 4, later this month, according to a report by The Information published on Friday.
The release, however, may be delayed once more, sources familiar with the matter told the publication.
The launch of Llama 4 has already been postponed at least twice, as Meta races to keep pace in the intensifying global competition for dominance in artificial intelligence. The tech giant, which owns Facebook, has faced challenges with Llama 4’s performance, particularly in technical benchmarks related to reasoning, mathematics, and humanlike voice conversation—areas where competitors like OpenAI have excelled.
Meta’s significant push into AI comes as industry leaders increase investment in machine learning technologies. The company plans to allocate as much as $65 billion this year toward expanding its AI infrastructure, under pressure from investors to show tangible returns.
The new model is expected to draw on certain innovations from DeepSeek, a rising Chinese AI firm that has produced a competitive, cost-effective alternative. Notably, one version of Llama 4 is said to adopt a “mixture of experts” approach—a machine learning method where different parts of the model are trained to specialize in distinct tasks, improving efficiency and accuracy.
Meta is reportedly considering a phased rollout strategy for Llama 4, initially integrating it into its Meta AI products before releasing it as open-source software.
Last year, the company introduced Llama 3, a largely free AI model that supported eight languages, delivered improved code generation, and solved more advanced math problems than its predecessors.
Meta has not officially confirmed the release date for Llama 4.
