Meta reportedly is postponing the release of the largest version of its open-source Llama 4 artificial intelligence (AI) model from summer to fall at the earliest.
Called “Behemoth,” the multimodal model is not improving “significantly” enough to be released by June; it was already delayed from April, when Meta held LlamaCon, its first Llama developers conference.
The delay looks to be the first hiccup from Meta on the release of its Llama flagship family of large language models, which have been praised for the speed of their release, according to The Wall Street Journal.
As a powerful open-source model, Llama has given developers in smaller companies, nonprofit communities and academia a generally free AI model to use. It has been the counterweight to the closed, proprietary models developed by OpenAI, Google, Amazon and others.
The impact on companies is more muted since many big companies go through the cloud giants, which mostly offer proprietary models.
Smaller companies can customize the smaller open-source Llama models, but still need help to implement them since Meta — as a social media giant — is not in the business of offering deployment services. Meta is using Llama to power its own social media tools, so CEO Mark Zuckerberg can control his own AI destiny.
The issue with Behemoth is whether the model shows enough advances to justify launching it publicly, according to the paper.
Need for Speed
In the tech industry, developers and users can quickly disparage new releases if they don’t show enough advances to justify a public launch.
At LlamaCon, Meta released two smaller sister Llama 4 models that are still large in certain aspects.
Maverick has 400 billion total parameters (internal settings) with a 1 million token context window length or 750,000 words (GPT-4o only has 128,000 tokens)
Scout has 109 billion parameters and a 10 million (7.5 million words) context window length.
Initially, Behemoth was set for release at the same time. It would have 2 trillion parameters.
The Journal said Meta is getting impatient with its Llama 4 team as it continues to pour a fortune into AI investments.
This year, the company has budgeted up to $72 billion in capital expenditures, much of which is earmarked for AI development in support of Zuckerberg’s long-term vision.
Read more: Meta Adds ‘Multimodal’ Models to Its Llama AI Stable
Mounting Frustrations
Zuckerberg and other senior leaders have yet to disclose a public release date for Behemoth. While the model could still launch earlier than expected, possibly in a limited form, insiders are worried that its current performance may not live up to expectations set by company statements.
Frustrations are reportedly mounting among Meta’s leadership over the progress made by the team responsible for the Llama 4 models, which has struggled to deliver tangible gains on Behemoth. This has led the company to consider major leadership changes in its AI product group.
Meta has publicly promoted Behemoth as a powerful system, claiming it surpasses offerings from OpenAI, Google and Anthropic on certain evaluations. Internally, however, training difficulties have hampered its effectiveness, people familiar with the development said.
PYMNTS contacted Meta for comment but has yet to get a reply.
OpenAI has also experienced delays. Its next major model, GPT-5, was originally anticipated for a mid-2024 release. Last December, the Journal noted that development had fallen behind schedule.
OpenAI CEO Sam Altman later clarified in February that the interim model would be GPT-4.5, while GPT-5, expected to deliver larger advances, remained months away.
Reasons for Delay
Advances in AI model development could slow for several reasons. Among them:
Running out of high-quality data
Large language models require massive amounts of data to train on, such as the entire internet. But they may be running out of publicly available data to access, while copyrighted content carries legal risks.
That is why OpenAI, Google and Microsoft are urging the Trump administration to preserve their right to train on copyrighted material.
“The federal government can both secure Americans’ freedom to learn from AI, and avoid forfeiting our AI lead to the PRC [People’s Republic of China] by preserving American AI models’ ability to learn from copyrighted material,” according to OpenAI.
Algorithmic limitations
It used to be that increasing model size, using more compute, and letting models train on more data would yield notable advances. But there have been diminishing returns from AI models, leading some to say the scaling laws are slowing down, according to Bloomberg.
