Meta has pushed back the release of its new A.I. model, code-named Avocado, to at least May after internal tests showed it trailing rivals Google, OpenAI and Anthropic.
Meta has delayed the release of its most ambitious artificial intelligence model after internal tests showed it falling short of competing systems from Google, OpenAI and Anthropic, three people with knowledge of the matter said — a setback for Mark Zuckerberg, who has staked hundreds of billions of dollars on his company’s ability to lead in the technology.
The model, known internally as Avocado, had been targeted for release this month. It has now been pushed back to at least May, said the people, who were not authorized to speak publicly about confidential matters. Meta’s A.I. leaders have also discussed temporarily licensing Google’s Gemini to power the company’s A.I. products in the interim, though no decisions have been reached.
Avocado outperformed Meta’s previous A.I. model and surpassed Google’s Gemini 2.5, released in March, on internal benchmarks for reasoning, coding and writing. But it has not matched Gemini 3.0, released by Google in November, the people said — leaving Meta’s flagship model a step behind the frontier it set out to define.
Zuckerberg, 41, said in July that Meta’s new A.I. models would “push the frontier in the next year or so.” Hitting that target now appears increasingly unlikely.
Also Read: Amazon’s Layoffs Aren’t a Warning. They’re a Preview.
Billions Spent, Ground Still to Make Up
The delay is the latest complication for a company that has made A.I. the organizing ambition of its next decade. Meta, which owns Facebook, Instagram and Threads, has committed $600 billion to building data centers to power the technology. In January, it projected spending as much as $135 billion this year alone — nearly twice the $72 billion it spent in 2024. The company has also hired aggressively, recruiting top A.I. researchers from across the industry.
Despite that investment, Google, OpenAI and Anthropic remain widely regarded as ahead in foundational A.I. models — the large, general-purpose systems that underpin chatbots, video generators, coding assistants and other products. Falling behind on foundational models carries consequences beyond any single product launch: it hampers a company’s ability to recruit leading researchers, slows internal experimentation and risks ceding the narrative in a field where perceived momentum matters as much as measured performance.
A spokesman for Meta, Dave Arnold, said in a statement: “Our next model will be good but, more importantly, will show the rapid trajectory we’re on. We’ll steadily push the frontier over the course of the year as we continue to release new models. We’re excited for people to see what we’ve been cooking very soon.”
A New Lab, a New Strategy — and Some Turbulence
Avocado was born from a broader strategic reset. After Meta’s previous model, Llama 4, fell short of expectations last year, Zuckerberg moved to accelerate the company’s A.I. ambitions on multiple fronts. In June, Meta invested $14.3 billion in Scale AI, a data and A.I. services start-up, and installed its chief executive, Alexandr Wang, 29, as Meta’s new chief A.I. officer. Zuckerberg declared that Meta’s goal was now to build a “superintelligent” form of A.I. that would usher in “a new era for humanity.”
 Wang assembled an elite internal research group called TBD Lab — shorthand for “to be determined” — which set to work on two new models with fruit-themed code names: Avocado, a foundational language model, and Mango, an image and video generator. TBD Lab, which employs around 100 people, completed the first stage of Avocado’s development, known as pre-training, at the end of last year. Post-training — the phase in which a model is refined for real-world use — began in January, when the team set a target release of mid-March.
That timeline has since slipped. A handful of researchers have departed TBD Lab ahead of the model’s release, and internal tensions have surfaced. Wang has clashed with Chris Cox, Meta’s chief product officer, and Andrew Bosworth, the chief technology officer, over how the new models should improve Meta’s advertising business — the engine that generates most of the company’s revenue.
Last week, Meta told employees it would create a new A.I. engineering team under Bosworth that would collaborate with Wang’s division, a reorganization first reported by The Wall Street Journal. The announcement prompted speculation that Zuckerberg and Wang were at odds. Meta moved quickly to tamp down the talk. A company spokesman called the suggestion “totally false,” and Zuckerberg posted a selfie with Wang on Threads with the caption: “Meanwhile at Meta HQ.”
Also Read: Cutting Through the Noise of SaaS Buying
Tempering Expectations, Planning Ahead
Zuckerberg began managing expectations for Avocado publicly in January, telling investors on an earnings call: “I expect our first models will be good, but more importantly, will show the rapid trajectory we’re on.”
The only product TBD Lab has released to date is Vibes, an A.I. video application similar to OpenAI’s Sora.
Meta’s leaders are already looking beyond Avocado. Discussions about the company’s next foundational model are underway, according to one person familiar with the matter. Its code name is Watermelon.
