Meta to postpone release of Llama 4 Behemoth model, report claims

Meta Platforms Inc. is likely to delay the release of its upcoming Llama 4 Behemoth artificial intelligence model, in a move that could have serious implications for the broader AI industry.

In an exclusive report, the Wall Street Journal cites people with knowledge of the matter as saying Llama 4 Behemoth’s debut is being pushed back from early summer to the fall of this year, or potentially even later. According to the report, Meta is struggling to improve the large language model’s capabilities enough to justify an earlier launch. It’s said to be worried that its performance won’t match earlier claims made by the company.

The report casts doubt on Meta’s multibillion-dollar AI strategy, and sent its stock down more than 3% during regular trading today.

The delay is a sign that development in the AI industry is potentially slowing, and raises questions about the vast amounts of money Meta has been spending on AI infrastructure. Earlier this year, Meta said it plans to spend up to $72 billion on capital expenditures this year, with most of that going toward AI. The Journal says internal frustration is mounting, as some executives reportedly blame the Llama 4 models team for the delayed progress on Behemoth.

Meta is said to be contemplating making some “significant management changes” to the AI product group responsible for Llama 4 Behemoth’s development, the Journal added. Chief Executive Mark Zuckerberg has not yet set a public timeline for the launch of Behemoth, and could yet decide to launch an earlier, more limited version of the new model.

The company had originally planned to launch Llama 4 Behemoth in April, in time for its first AI developer conference, but it later pushed the date back to June. Now, the timeline is much less certain, with the Journal indicating that Meta’s AI engineering and research teams have serious concerns that the new model won’t live up to earlier claims about its capabilities.

This isn’t the first time reports have surfaced about Meta’s struggles with recent Llama models. The Information has also reported on problems at the company, and Meta itself has acknowledged submitting a specially optimized version of Llama to a leaderboard in April, rather than the publicly available version.

One of the company’s senior AI engineers, Ahmad Al-Dahle, admitted in a post on X last month that the company is hearing “reports of mixed quality across different services.”

We’re glad to start getting Llama 4 in all your hands. We’re already hearing lots of great results people are getting with these models.

That said, we’re also hearing some reports of mixed quality across different services. Since we dropped the models as soon as they were…

— Ahmad Al-Dahle (@Ahmad_Al_Dahle) April 7, 2025

The delay is embarrassing for Meta because the company has previously made some bold claims, saying that Llama 4 Behemoth outperforms the likes of GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 Pro on key benchmarks such as MATH-500 and GPQA Diamond, even while it’s still being trained.

That said, it seems Meta is not alone in its struggles. A similar situation appears to have arisen at OpenAI, which initially aimed to launch GPT-5 by the middle of this year, before ultimately releasing that model as GPT-4.5. The GPT-5 name has now been given to an upcoming “reasoning” model that’s still in development. In February, OpenAI CEO Sam Altman stated that any major new breakthroughs are still months away.

Another AI company, Anthropic PBC, delayed the release of its long-awaited Claude 3.5 Opus model, which has not yet been launched despite earlier claims it was “coming soon.”

Experts told the Journal that the apparent stumbles indicate that any advancements in AI are likely to come at a slower pace in future, and potentially cost much more in terms of investment. “The progress is quite small across all the labs, all the models,” said Ravid Shwartz-Ziv, an assistant professor at New York University’s Center for Data Science.

Meta’s struggles aren’t helped by the fact that most of the researchers who helped to create the original Llama model that debuted in early 2023 are no longer with the company. The original Llama team consisted of 14 academics and researchers with doctorate degrees, but 11 of them have since left the company. More recent versions of Llama have been developed by an entirely different team, the Journal said.

Image: SiliconANGLE/Dreamina