Meta delays flagship AI model amid internal doubts and leadership shake-ups

Meta delays flagship AI model

Meta is delaying the release of its most ambitious large language model, codenamed Behemoth, in what insiders describe as a moment of reckoning for the company’s massive artificial intelligence investments. Once internally slated for a splashy launch in April to coincide with its first AI developer conference, the model has now been pushed back to fall or later, according to people familiar with the matter, the Wall Street Journal reported.

The postponement comes amid internal concerns that Behemoth’s performance does not significantly surpass earlier versions of Meta’s Llama models—raising questions about whether the product is ready for public release. Some employees fear that releasing the model too soon could undercut the company’s public claims about its technological prowess in the ongoing AI arms race.

Mounting pressure and missed benchmarks

Meta has spent tens of billions of dollars in its pursuit of AI dominance, planning up to $72 billion in capital expenditures this year alone. Much of that investment is aimed at powering generative AI features across its social media platforms, including Facebook, Instagram, and WhatsApp.

Despite public boasts that Behemoth outperforms competitors like OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude models on certain benchmarks, insiders say development has been hampered by persistent training challenges. Engineers working on the model remain unconvinced that it delivers the kind of breakthrough Meta has promised—and some are worried that its current form doesn’t meet internal expectations or external hype.

Senior executives are reportedly frustrated with the Llama 4 team, which has been overseeing Behemoth’s development. Sources familiar with internal discussions say Meta is considering major leadership changes within its AI product group as a result.

A credibility test for Meta’s AI strategy

The delay has intensified scrutiny over Meta’s AI roadmap, especially after revelations that the company submitted a customized version of its model to a prominent leaderboard test—an optimized variant that differed from what was ultimately released to the public. While the submission briefly earned Meta top-tier rankings, the disclosure raised ethical concerns and prompted clarification from the leaderboard’s administrators. CEO Mark Zuckerberg later admitted the benchmark model was tailored specifically for that test.

Meta’s AI credibility is further challenged by a brain drain from its original Fundamental AI Research Team. Of the 14 researchers who co-authored Meta’s original Llama model paper in early 2023, 11 have since left the company.

Meta not alone in AI growing pains

Meta’s setbacks come as rivals also face headwinds. OpenAI’s much-anticipated GPT-5 has been delayed, with CEO Sam Altman confirming that only a 4.5 version would arrive this year. Anthropic’s Claude 3.5 Opus, a larger and more powerful model than previous releases, has yet to debut. These delays suggest that generative AI breakthroughs are becoming harder and more expensive to achieve, despite investor and public pressure to show rapid progress.

“Right now, the progress is quite small across all the labs, all the models,” said Ravid Shwartz-Ziv, a faculty fellow at NYU’s Center for Data Science.

No timeline, but billions on the line

Meta has not committed publicly to a release date for Behemoth. A company spokesperson declined to comment. Some insiders say the company may opt to release a limited or scaled-down version of the model to maintain momentum without fully committing to a model that falls short of expectations.

As the AI arms race evolves, Meta’s delay underscores a broader reality facing the sector: that developing truly next-generation AI is proving more complex, more expensive, and slower than anticipated. And for a company staking its future on AI, the pressure to deliver remains immense.