Meta Delays Release of Behemoth AI Model
Meta has postponed the release of its largest Llama 4 artificial intelligence model, codenamed “Behemoth,” until fall or later. The delay is attributed to the model’s lack of significant improvements, according to a report by The Wall Street Journal.
The Behemoth model, which was initially expected to be released in April alongside two smaller Llama 4 models, Maverick and Scout, has been delayed due to its failure to show substantial advancements. The model was anticipated to have 2 trillion parameters, significantly more than its smaller counterparts, with Maverick having 400 billion parameters and Scout having 109 billion parameters.
The delay in releasing Behemoth may not have a significant impact on companies, as they already have access to other open-source Llama models. Many large companies rely on cloud giants that offer proprietary models, while smaller companies can customize smaller open-source Llama models, although they require assistance with implementation.
Meta’s decision to delay the release of Behemoth highlights the challenges faced by the company in developing its AI models. The company has invested heavily in AI development, with a budget of up to $72 billion in capital expenditures this year, a significant portion of which is earmarked for AI research.
The delay is not unique to Meta, as other companies, such as OpenAI, have also experienced setbacks in their AI development. OpenAI’s GPT-5 model, for instance, was initially expected to be released in mid-2024 but has been delayed.
Experts point to several reasons for the slowdown in AI advancements, including the lack of high-quality training data and algorithmic limitations. The industry is facing diminishing returns from increasing model size and training data, leading some to suggest that the scaling laws are slowing down.
As the AI landscape continues to evolve, companies like Meta are under pressure to deliver innovative solutions while navigating the challenges associated with developing complex AI models.
