Mark Zuckerberg’s Meta is reportedly delaying the release of its next artificial intelligence model at least about two months after internal tests showed the tech falling short of rival models from Google, OpenAI and Anthropic.
The leaders of Meta’s AI division are considering temporarily licensing Google’s Gemini to power its products in the wake of the lackluster results, the New York Times reported this week.
Internal Meta testing showed lackluster performance for its new “Avocado” AI model even after techies worked on it for months, anonymous sources told the Times.
Instead of rolling out this month, Avocado is reportedly expected to launch around May after it failed to measure up in terms of reasoning, coding and writing.
The setbacks come after Zuckerberg, 41, threw the kitchen sink at developing AI. Meta has spent billions hiring top AI talent and committed $600 billion to building data centers to power the tech. Meta projected in January it would spend as much as $135 billion this year alone — nearly twice the $72 billion it blew through last year developing AI.
Zuckerberg said in July that Meta’s new AI models would “push the frontier in the next year or so.”
The Times noted that Avocado outperformed Meta’s previous AI model and did better than Google’s Gemini 2.5 from March — but lagged behind Gemini 3.0 from November.
“I expect our first models will be good, but more importantly will show the rapid trajectory we’re on,” Zuckerberg said on an investor call in January.
A Meta spokesperson said: “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.”
AI efforts by Meta, the owner of Facebook, Instagram and Threads, are being closely watched in the artificial intelligence arms race. Google, OpenAI and Anthropic are widely regarded as ahead in developing foundational AI models — the underlying technology for chatbots, video generators and coding tools.
Zuckerberg kicked his AI effort into overdrive after Meta’s previous model, Llama 4, fell short of expectations last year. To catch up, the company invested $14.3 billion in the startup Scale AI in June and made its 29-year-old chief executive, Alexandr Wang, its new chief AI officer.
Wang helped assemble an AI lab within Meta called TBD Lab. The unit finished the first stage of Avocado’s development, called “pre-training,” at the end of last year, according to the Times. In January, it began the next phase, “post-training,” and set a target release date of mid-March.
The only product Wang’s AI division has released so far is Vibes, an AI video app similar to OpenAI’s Sora.
TBD Lab, which has around 100 employees, has seen some turnover, with a handful of researchers departing before Avocado’s release, according to the report.
Meta’s top brass have reportedly debated whether or not its new AI model will be “open source,” meaning parts of its code would be available to other developers.
Meta has long championed open source models, arguing that they help advance tech, while companies like OpenAI and Anthropic have said letting others build off their AI would pose safety risks.
The Times article detailed tensions between Wang and Chris Cox, Meta’s chief product officer, and Andrew Bosworth, the chief technology officer. Disagreements reportedly centered on how new AI models should improve Meta’s ad business.
Meta wrote in a note to employees last week, which was reported earlier by the Wall Street Journal, that it would create an AI team under Bosworth that would work with Wang.
Last week, Meta told employees it was creating an AI team under Bosworth that would work with Wang, according to the Wall Street Journal.
