From Llamas to Avocados: Meta's shifting AI strategy is causing internal confusion

Meta CEO Mark Zuckerberg was truthful optimistic past twelvemonth astir his company's Llama household of artificial quality models that helium predicted they would go the "most precocious successful the industry" and "bring the benefits of AI to everyone."

But aft including a full conception connected Llama successful his opening remarks during Meta's earnings call successful January of this year, helium mentioned the marque sanction lone erstwhile connected the latest telephone successful October. The company's obsession with its open-source ample connection exemplary has fixed mode to a precise antithetic attack to AI, 1 focused astir a multibillion-dollar hiring spree to bring successful apical manufacture endowment that could assistance Meta instrumentality connected the likes of OpenAI, Google and Anthropic.

As 2025 comes to a close, Meta's strategy remains scattershot, according to insiders and manufacture experts, feeding the cognition that the institution has fallen further down its apical AI rivals, whose models are rapidly gaining adoption successful the user and endeavor markets.

Meta is pursuing a caller Llama successor and frontier AI model, codenamed Avocado, CNBC has learned. People with cognition of the substance said galore wrong the institution were expecting the exemplary to beryllium released earlier the extremity of this year, but that the program present is for that to hap successful the archetypal 4th of 2026. The exemplary is wrestling with assorted training-related show investigating intended to guarantee the strategy is well-received erstwhile it yet debuts, said the people, who asked not to beryllium named due to the fact that they weren't authorized to talk connected the matter.

"Our exemplary grooming efforts are going according to program and person had nary meaningful timing changes," a Meta spokesperson said successful a statement.

With its banal underperforming the broader tech assemblage this twelvemonth and severely trailing Google genitor Alphabet, Wall Street is looking for a sense of direction and a way to a instrumentality connected concern aft Meta spent $14.3 cardinal successful June to prosecute Scale AI laminitis Alexandr Wang and a fistful of his apical engineers and researchers. Four months aft that announcement, which included Meta purchasing a large involvement successful Scale, the societal media institution raised its 2025 guidance for capital expenditures to betwixt $70 cardinal and $72 cardinal from a anterior scope of $66 cardinal to $72 billion.

"In galore ways, Meta has been the other of Alphabet, wherever it entered the twelvemonth arsenic an AI victor and present faces much questions astir concern levels and ROI," analysts astatine KeyBanc Capital Markets wrote successful a November enactment to clients. The steadfast recommends buying some stocks.

At the bosom of Meta's situation is the sustained dominance of its halfway business: digital advertising.

Even with yearly income successful excess of $160 billion, Meta's advertisement targeting business, driven by monolithic improvements successful AI and the popularity of Instagram, is increasing gross northbound of 20% a year. Investors person lauded the institution for utilizing AI to bolster the strength of its currency cattle and to marque the enactment much businesslike and little bloated.

But Zuckerberg has overmuch grander ambitions, and the caller defender he's brought successful to propulsion the aboriginal vision of AI has nary inheritance successful online ads. The 41-year-old founder, with a nett worthy of implicit $230 billion, has suggested that if Meta doesn't instrumentality large swings, it risks becoming an afterthought successful a satellite that's poised to beryllium defined by AI.

Until recently, Meta's unique position successful AI was the open-source quality of its Llama models. Unlike different AI models, Meta's exertion was made freely disposable truthful third-party researchers and others could entree the tools and yet amended them.

"Today, respective tech companies are processing starring closed models," Zuckerberg wrote successful a blog station successful July 2024. "But unfastened root is rapidly closing the gap."

He's since started changing his tune. Zuckerberg hinted implicit the summertime that Meta was considering shaking up its attack to unfastened root aft the April merchandise of Llama 4, which failed to captivate developers. Zuckerberg said successful July that, "We'll request to beryllium rigorous astir mitigating these risks and cautious astir what we take to unfastened source."

Avocado, erstwhile it's yet made available, could beryllium a proprietary model, according to radical acquainted with the matter. That means extracurricular developers wouldn't beryllium capable to freely download its alleged weights and related bundle components. 

Some astatine Meta were upset that the R1 model released by Chinese AI laboratory DeepSeek earlier this twelvemonth incorporated pieces of Llama's architecture, the radical said, further underscoring the risks of unfastened root and hammering location the thought that the institution should overhaul its strategy.

The company's high-priced AI hires and leaders of the precocious launched Meta Superintelligence Labs, oregon MSL, person besides questioned the open-source AI strategy and favored creating a much powerful proprietary AI model, CNBC reported successful July. A Meta spokesperson said astatine the clip that the company's "position connected unfastened root AI is unchanged."

The Llama 4 flub was a important catalyst successful Zuckerberg's pivot, the radical said, and besides led to a large interior shakeup. Chris Cox, Meta's main merchandise serviceman and a 20-year institution seasoned who was hired arsenic its 13th bundle engineer, nary longer oversees the AI division, formally known arsenic the GenAI unit, aft the botched release, the radical said.

Zuckerberg went connected a spending spree to retool Meta's AI leadership.

He landed connected Wang, past Scale AI's 28-year-old CEO, who was named Meta's caller main AI serviceman and, successful August, became the caput of an elite portion called TBD Lab. Avocado is being developed wrong TBD, radical acquainted with the substance said.

OpenAI CEO Sam Altman said successful June that Meta was trying to lure endowment from his institution with gigantic wage packages, including sky-high $100 cardinal signing bonuses, which Meta said astatine the clip was a misrepresentation.

Along with Wang came different tech bigwigs, including erstwhile GitHub CEO Nat Friedman, who heads the merchandise and applied probe limb of MSL, and Shengjia Zhao, who was a ChatGPT co-creator. They've brought with them modern methods that person go the modular for frontier AI improvement successful Silicon Valley, and person upended the accepted bundle improvement process wrong Meta, the radical said.

Wang is present nether unit to present a top-tier AI exemplary that helps the institution regain momentum against rivals similar OpenAI, Anthropic and Google, the radical said. 

That unit has lone accrued arsenic competitors stepped up their game. Google's Gemini 3, unveiled past month, has drawn coagulated reviews from users and analysts. OpenAI precocious announced caller updates to its GPT-5 AI model, portion Anthropic debuted its Claude Opus 4.5 model in November soon aft releasing 2 different major models.

Analysts antecedently told CNBC that there's nary wide starring AI model, due to the fact that immoderate execute amended connected definite tasks similar conversations oregon coding. But the 1 changeless is that each of the large exemplary creators person to walk a batch of money connected AI to support immoderate competitory edge, they said.

A hefty dose of that spending lines the pockets of Nvidia, the starring developer of AI graphics processing units. Nvidia CEO Jensen Huang laid retired the authorities of play during his company's earnings call successful November, aft the chipmaker reported 62% year-over-year gross growth. He highlighted a fig of exemplary developers arsenic large customers, including Elon Musk's xAI.

"We tally OpenAI. We tally Anthropic. We tally xAI due to the fact that of our heavy concern with Elon and xAI," Huang said. "We tally Gemini. We tally Thinking Machines. Let's see, what other bash we run? We tally them all."

At nary constituent did Huang notation Llama, though elsewhere connected the telephone helium said Meta's Gem, "a instauration exemplary for advertisement recommendations trained connected large-scale GPU clusters," drove an betterment successful advertisement conversions astatine Meta successful the 2nd quarter.

Wang isn't the lone Meta exec feeling the heat.

Friedman has besides been tasked with producing a breakout AI product, the radical said. He was liable for Meta's September motorboat of Vibes, a provender of AI-generated abbreviated videos, which is wide viewed arsenic inferior to OpenAI's Sora 2, they said. Former employees and creators told CNBC that the merchandise was rushed to marketplace and lacked cardinal features, similar the quality to make realistic lip-synced audio.

Although Vibes has attracted much involvement to the company's standalone Meta AI app, it trails the Sora app arsenic measured by downloads, according to information provided to CNBC by Appfigures.

Pressure is being felt crossed Meta's AI organizations, wherever 70-hour workweeks person go the norm, the radical said, portion teams person besides been deed with layoffs and restructurings passim the year.

In October, Meta chopped 600 jobs successful MSL to trim layers and run much quickly. Those layoffs impacted employees successful areas similar the Fundamental Artificial Intelligence Research unit, oregon FAIR, and played a cardinal relation successful Chief AI Scientist Yann LeCun's determination to permission the institution to motorboat a startup, according to radical with cognition of the matter.

LeCun declined to comment.

Zuckerberg's high-stakes determination to crook to outsiders similar Wang and Friedman to pb the company's AI efforts represented a large alteration for a institution that's historically promoted long-tenured workers to apical posts, the radical said.

In Wang and Friedman, Zuckerberg has handed the controls to experts successful infrastructure and related systems, alternatively than user apps. The caller leaders besides brought a antithetic absorption benignant and 1 that's unfamiliar wrong Meta.

In particular, insiders told CNBC that Wang and Friedman are much cloistered successful their communications, a opposition to a historically unfastened attack of sharing enactment and chatting connected the company's Workplace interior societal web

Members of Wang's TBD Lab, who enactment adjacent Zuckerberg's office, don't usage Workplace, radical acquainted said, adding that they're not adjacent connected the web and that the radical operates similar a abstracted startup.

However, Zuckerberg isn't giving the caller AI enactment squad implicit autonomy. Engineering vice president Aparna Ramani, who has been with Meta for astir a decade, has been enactment successful complaint of overseeing the organisation of computing resources for MSL, the radical said.

And successful October, Vishal Shah was moved from starring the company's metaverse initiatives wrong Reality Labs, wherever he'd been for 4 years, to a caller relation arsenic vice president of AI Products, moving with Friedman. Shah is considered a loyal lieutenant who has helped enactment arsenic a span betwixt the company's accepted societal apps similar Instagram and newer projects similar Reality Labs, the radical said.

Meta confirmed past week that it plans to chopped resources to its virtual world and related metaverse initiatives, shifting its attraction to its fashionable AI-infused glasses developed with EssilorLuxottica.

One of the biggest points of hostility betwixt the aged and the caller is successful the realm of bundle development, radical acquainted with the substance said.

In creating products, Meta has traditionally sought input from galore groups liable for areas similar front-end idiosyncratic interface, design, algorithmic feeds and privacy, the radical said. The multi-step process was intended to guarantee immoderate level of uniformity among the company's apps that pull billions of users each day.

But the galore interior tools built implicit the years to assistance coders make bundle and features weren't developed to accommodate instauration models. Meta's caller AI leaders, notably Friedman, presumption them arsenic bottlenecks slowing down what should beryllium a rapid-fire improvement process, the radical said.

Friedman has called for MSL to usage newer tools that person been calibrated to incorporated aggregate AI models and assorted kinds of coding automation bundle often called AI agents, the radical said.

"They person this mantra present saying 'Demo, don't memo,'" Lovable CEO Anton Osika said successful October astatine the Masters of Scale Summit successful San Francisco, astir Meta's caller improvement process.

Osika said Meta employees person been utilizing Lovable's tools to much rapidly physique interior apps, specifically referencing the company's concern teams, which person turned to Lovable to make bundle for tracking headcount and planning.

While Meta continues retooling its app improvement methods and pushes toward releasing Avocado, the institution has been experimenting with different AI models connected its products. Vibes, for instance, relied connected AI models from Black Forest Labs and Midjourney, a startup that counts Friedman arsenic an advisor.

Meta is besides altering its attack to infrastructure, and is progressively turning to third-party unreality computing services similar CoreWeave and Oracle for processing and investigating AI features arsenic it builds retired its ain monolithic information centers, the radical said.

The societal media elephantine announced successful October that it signed a associated task statement with Blue Owl Capital arsenic portion of a $27 cardinal woody to assistance money and make the gargantuan Hyperion information halfway successful Richland Parish, Louisiana. The institution said astatine the clip that the concern provides the "the velocity and flexibility" Meta needs to physique the information halfway and enactment its "long-term AI ambitions."

Despite the company's challenges successful 2025, Zuckerberg's connection to employees and investors is that he's much committed than ever to winning. At the apical of the company's net telephone successful October, Zuckerberg said MSL is "off to a beardown start."

"I think that we've already built the lab with the highest talent density in the industry," Zuckerberg said. "We're heads down processing our adjacent procreation of models and products and I'm looking guardant to sharing much connected that beforehand implicit the coming months."

WATCH: Data halfway request driving infrastructure credit