Meta’s Behemoth AI Model Slips to Fall, Stock Drops 2.4%

Meta delayed its flagship Behemoth AI model to fall 2025 or later, sending shares of the social media company down 2.4% on the news. The Wall Street Journal reported that engineers at Meta are struggling to significantly improve the model’s capabilities, prompting internal questions about whether the gains are large enough to justify a public release. Behemoth was originally scheduled to debut in April at LlamaCon 2025, Meta’s first developer conference focused on its open-source large language models, before slipping to June and then to its new fall-or-later window.

The single-day move is small by Meta’s standards, but the Journal’s report lands on a balance sheet built around a much larger AI footprint. Meta has guided to $60 billion to $72 billion in capital expenditure for 2025, a figure the company raised earlier this year as it expanded its data center plans. Behind the modest slide, the report points to senior executives openly frustrated with the team that built the smaller Llama 4 models, and considering management changes to the AI product group.

Behemoth’s Repeated Delays

Meta’s flagship AI model has now been pushed back twice. According to the original report on Meta’s Behemoth delay, Behemoth was originally scheduled to debut in April 2025 in alignment with LlamaCon 2025, the company’s first developer conference focused on its open-source large language models.

That April window slipped, with the release first pushed to June, then to fall or possibly later, the Journal reported, citing people familiar with the matter. The progression matters because Behemoth is described by Meta as a teacher model, a much larger system designed to distil its capabilities into smaller, more deployable open-weight models like Llama 4 Scout and Llama 4 Maverick, both of which were released in April. Without Behemoth shipping on schedule, the smaller models lose the upstream system that was meant to keep them competitive. Each postponement pushes that demonstration further into a calendar that already has competitors shipping faster-reasoning systems on tighter cycles.

The repeated delay also tests the timing of Meta’s open-source strategy. Behemoth was positioned internally as the proof point that Meta’s bet on open-weight models could keep pace with the closed systems built by OpenAI, Anthropic, and Google.

1. April 2025: Behemoth originally slated to debut at LlamaCon 2025; Meta releases Llama 4 Scout and Llama 4 Maverick.
2. May 2025: Release pushed from June to fall or later, per the Wall Street Journal.
3. Fall 2025 or later: New target window for the Behemoth release.

Why the Engineers Are Stuck

The reason for the delay is not scheduling. According to the Journal, Meta engineers are struggling to significantly improve the capabilities of the Behemoth large-language model, leading to staff questions about whether the gains over prior versions are large enough to justify a public release. The pattern fits a broader industry phenomenon: early generations of generative-AI models delivered exponential leaps in performance, and newer large models are not matching that curve. The performance gap between what Behemoth can deliver and what the smaller Llama 4 models already ship is reportedly not as wide as the company had planned. That gap is the core technical problem holding up the release.

The result is an internal debate over whether to ship the model at all. Engineers and researchers are reportedly concerned that Behemoth’s performance may not live up to the company’s public claims about its capabilities, the Journal reported. That uncertainty is feeding a broader anxiety inside Meta’s AI product group about how to position the model when the smaller Llama 4 variants have already received a tepid reception from developers who compared them unfavourably to DeepSeek and Alibaba’s Qwen.

• Llama 4 Scout: Released April 5, 2025 as a 17B active parameter model with 16 experts, designed to run on a single NVIDIA H100 GPU with on-the-fly int4 quantization.
• Llama 4 Maverick: Released April 5, 2025 as a 17B active parameter model with 128 experts, positioned against other foundation models in Meta’s open-weight lineup.
• Llama 4 Behemoth: Not yet released; a teacher model designed for distillation into smaller, specialised open-weight systems. See the Llama 4 Scout model release page for the family architecture.

Senior Executives Turn on the Llama 4 Team

The engineering trouble has now reached the executive suite at Meta. The Journal reports that senior leaders are openly blaming the team that built the smaller Llama 4 models for the failure to push Behemoth across the line.

Senior executives at the company are frustrated at the performance of the team that built the Llama 4 models and blame them for the failure to make progress on Behemoth, according to people familiar with their views. Meta is contemplating significant management changes to its AI product group as a result, the people said.

Citing people familiar with the executives’ views, the Journal reports that the contemplated changes reach into the AI product group itself, the organisation responsible for both the smaller Scout and Maverick models and the larger Behemoth effort. Meta is reviewing the leadership of the group, a sign that the delay has moved from a technical problem to an organisational one. The frustration is reportedly being expressed privately by senior leaders, not on the record, which is why the people familiar with their views are the only public window into the dispute. That same dynamic, internal critique without public acknowledgment, has been a recurring pattern across Meta’s AI efforts over the past year.

The public reception at LlamaCon 2025 in April hinted at the underlying tension. Several developers told Business Insider at the conference that they had expected a reasoning model to be announced at the inaugural event and would have settled for a traditional model that could beat alternatives such as DeepSeek’s V3 and Alibaba’s Qwen. Vineeth Sai Varikuntla, a developer working on medical AI applications, told the publication that “Qwen is ahead, way ahead of what they are doing in general use cases and reasoning.”

Meta declined to comment on the Journal’s report on the contemplated management changes. A Meta spokesperson, Ryan Daniels, told Business Insider the company was “constantly listening to feedback from the developer community and our partners to make our models and features even better.”

A $72 Billion Bet Waiting on Behemoth

The delay lands on a balance sheet that has been built around a much larger AI footprint. Meta has guided to $60 billion to $72 billion in capital expenditure for 2025, a figure the company raised from an earlier range as it expanded its data center plans. The largest of those projects is the Richland Parish Data Center in Louisiana.

Richland Parish is a more than $10 billion investment that Meta says will deliver over two gigawatts of compute capacity, sized to train future open-source large language models, according to the company’s Meta’s Richland Parish Data Center project page. The site, and the rest of Meta’s 2025 build-out, was planned on the assumption that Behemoth and the rest of the Llama 4 family would be running at scale. The smaller Scout and Maverick models, released in April, can use that capacity. A Behemoth release that slips by a quarter or more pushes the revenue case for the new infrastructure into 2026 and beyond.

That timing matters because the capex bill is hitting at the same moment Meta’s competitors are signalling caution. Microsoft has reportedly pulled back on some of its large commitments for AI data centers, the Journal reported, even as DeepSeek and other smaller-model builders have demonstrated that they can match or beat the scores of some of the larger and more expensive systems. For Meta, the math on a $72 billion spend gets harder to defend if the model that anchors the spend is delayed and the wider industry is starting to question whether the largest models are still the right bet.

Why the Market Barely Blinked

The stock market’s reaction was relatively contained. Meta shares closed down 2.4% on the day of the Journal’s report, a fraction of a percent off the broader market, which edged higher. For a company whose stock had been trading near record highs, a single-day move of that size is a rounding error, and several market watchers quickly framed the delay as a calculated move rather than a structural setback. The argument is that releasing a half-finished flagship would do more damage to user trust and developer mindshare than a clean delay.

• $60-72B: Meta’s guided 2025 capital expenditure range, raised earlier this year.
• $10B+: Meta’s investment in the Richland Parish Data Center in Louisiana.
• 2+ gigawatts: Compute capacity the Richland Parish site is designed to deliver.
• April 5, 2025: Release date for Llama 4 Scout and Llama 4 Maverick, the two smaller open-weight models.

Meta’s core advertising business, anchored on Facebook, Instagram, and WhatsApp, continues to generate the cash flow needed to fund the AI build-out, and the company has said its superintelligence push will be measured in years, not quarters. The market’s tepid reaction suggests investors are, for now, willing to give Meta the time the company says it needs. The harder test comes when Meta next reports earnings, and management has to explain to shareholders why a flagship model that was meant to anchor a $72 billion spend is not yet ready to ship.

What the Delay Means for Open-Source AI

The Behemoth delay also tests Meta’s open-source credibility at a moment when that credibility is already under pressure. The company built its AI brand on the back of Llama 2 and Llama 3, the open-weight models that Nvidia CEO Jensen Huang once called “probably the biggest event in AI” in 2023, and on PyTorch, the machine learning framework Meta created in 2016 and transferred to the Linux Foundation in 2022. Behemoth was meant to extend that pattern into the largest, most expensive class of models.

That credibility took a hit in April, when the release of Llama 4 drew accusations that Meta had gamed the leaderboards by using a different version of the model for public benchmarking than the one available for download. The company denied the charge, saying the variant in question was experimental and that evaluating multiple versions of a model is standard practice. The Behemoth delay, on top of that, hands critics a fresh opportunity to argue that the gap between Meta’s marketing and its shipped models is widening. The competitive picture is also tightening. DeepSeek’s V3 and Alibaba’s Qwen are reportedly matching or beating the scores of Meta’s Llama models and OpenAI’s o1 on several benchmarks, and Microsoft’s pullback on data center commitments suggests even the largest players are questioning the size of the bet. For an open-source community that has come to expect Meta to ship the next big model, the Behemoth delay is a reminder that even the most ambitious open-weight roadmaps are now running into the same performance wall the closed labs are hitting.

Meta faces a separate fight on the regulatory front as well: in a related case, the EU order reopening WhatsApp to rival AI chatbots is forcing the company to restore third-party access to its messaging platform, a constraint that shapes which AI products Meta can ship to European users. The two fights are different in kind, but they share a theme. Meta’s AI roadmap is being contested on every front at once.

Frequently Asked Questions

What is Meta’s Behemoth AI model?

Behemoth is the largest model in Meta’s Llama 4 family, previewed by the company in April 2025 as a “teacher model” designed to distil its capabilities into smaller, more deployable open-weight systems like Llama 4 Scout and Llama 4 Maverick. It has not yet been released to developers.

When was Behemoth originally supposed to launch?

Behemoth was first scheduled for April 2025 at LlamaCon, Meta’s first developer conference dedicated to its open-source large language models. The release slipped to June, then to fall 2025 or later, according to the Wall Street Journal.

Why did Meta delay Behemoth?

Engineers are struggling to significantly improve the model’s capabilities over prior versions, the Journal reported, citing people familiar with the matter. The slowdown fits an industry pattern in which the largest models are no longer delivering the exponential leaps in performance that earlier generations produced.

How did Meta’s stock react to the delay?

Meta shares closed down 2.4% on the day of the Journal’s report, according to the Sherwood News write-up. The slide came as some tech giants, including Microsoft, were already pulling back on large AI data center commitments.