Meta has pushed back the release of its next foundational AI model, code-named Avocado, from this month to at least May 2026, the New York Times reported Thursday, March 12, citing three people with knowledge of the matter.
The delay follows internal tests in which Avocado fell short of the performance benchmarks set by leading AI models from Google, OpenAI, and Anthropic across reasoning, coding, and writing tasks, according to the Times report.
The model did outperform Meta's previous AI model and cleared the bar set by Google's Gemini 2.5, which launched in March 2025. But it did not reach the level of Google's Gemini 3.0, released in November 2025, placing Avocado between two generations of Google's most advanced models on Meta's internal evaluations.
Meta spokesman Dave Arnold responded to the Times in a statement saying, "our next model will be good, but more importantly, show the rapid trajectory we're on, and then 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."
Meta shares dropped 1% in premarket trading on Friday to $632.15 following the report.
Damilare Odedina
What Avocado Is and Who Built It
Multiple reports have portrayed Avocado as a foundational AI model, the type of large-scale system that powers chatbots, content generation tools, coding assistants, and recommendation products. According to the leaks, it is not expected to be a standalone consumer app. It will be the underlying engine Meta intends to deploy across its platforms.
The model is most likely being developed inside TBD Lab, an internal AI division Meta built after appointing Alexandr Wang, 29, as its Chief AI Officer. Wang's role came after Meta invested $14.3 billion in his company, Scale AI, in June 2025. TBD Lab has roughly 100 employees and has been working on two models simultaneously, Avocado for text and reasoning, and Mango for image and video generation.
Pre-training on Avocado, reports have claimed, finished at the end of last year. Post-training, the phase that shapes how a model responds to instructions, reportedly began in January 2026, when the team set a mid-March target for release. The only product TBD Lab has shipped so far is Vibes, an AI video app.
In a February memo viewed by The Information, Meta described Avocado internally as its "most capable pre-trained base model to date."
The Gemini Licensing Discussion and What Comes Next
The Times also reported that leaders within Meta's AI division discussed the possibility of temporarily licensing Google's Gemini to run Meta's AI products while Avocado continues development. The Times reported that no decisions have been reached on that option.
The Times reported internal disagreements between Wang and Meta's Chief Product Officer Chris Cox and Chief Technology Officer Andrew Bosworth over how the new AI models should support Meta's advertising business. In response, Meta announced internally last week that it would create a new AI engineering team under Bosworth to work alongside Wang's division, a move earlier reported by the Wall Street Journal.
After speculation that Zuckerberg and Wang had a falling out, Meta called the reports "totally false." Zuckerberg then posted a photo with Wang on Threads with the caption "Meanwhile at Meta HQ."
On the open-source question, whether Avocado's code will be made publicly available for developers or kept closed, no final decision has been reported. The Times noted that by last summer, Zuckerberg and Wang were leaning toward a closed model, a shift from Meta's historically open-source stance.
Meta's next model in the pipeline after Avocado has already been named Watermelon.
Damilare Odedina
