Meta's Next-Generation AI Model Faces Significant Delay After Benchmark Struggles
Meta's highly anticipated next-generation artificial intelligence model, internally codenamed Avocado, has encountered a substantial setback. According to a report from the New York Times published on Thursday, the model's release has been postponed by at least two months after it failed to meet performance expectations in internal benchmark comparisons against competing AI systems.
From March to May: A Critical Timeline Shift
Originally targeted for a mid-March deployment, the Avocado model is now unlikely to ship before May, according to three individuals with direct knowledge of the development timeline who spoke with the publication. This delay represents a significant stumble for Meta, a company that has invested billions of dollars in an aggressive effort to catch up in the foundational AI race dominated by competitors like Google, OpenAI, and Anthropic.
Benchmark Performance: Behind the Frontier Despite Massive Investment
While Avocado reportedly demonstrated improvements over Meta's previous Llama 4 model and even edged past Google's older Gemini 2.5, it could not keep pace with Google's more advanced Gemini 3.0, which was released in November. This performance gap places Meta's offering squarely behind the current frontier of AI capabilities, a concerning position given the company's monumental financial commitment.
Meta has allocated up to $135 billion in capital spending for this year alone—nearly double the $72 billion it expended in 2024—with a significant portion directed toward AI infrastructure and development. The benchmark shortfall raises fundamental questions about whether sheer financial firepower alone can bridge the technological gap with established AI leaders.
Internal Discussions Reveal Urgent Need for Competitive Edge
Perhaps the most revealing detail from the New York Times report concerns internal discussions among Meta's AI leadership. Executives have reportedly floated the idea of temporarily licensing Google's Gemini technology to power Meta's consumer products while the Avocado model undergoes further refinement. Although no final decision has been made, the mere consideration of this option underscores the pressing urgency Meta faces in deploying a competitive AI model to the market.
The TBD Lab Team and Internal Challenges
The Avocado model is being developed by TBD Lab, an elite research unit comprising approximately 100 personnel assembled under the leadership of Alexandr Wang. Wang, the former CEO of Scale AI, joined Meta as chief AI officer last June as part of a substantial $14.3 billion investment initiative. TBD Lab completed the pre-training phase for Avocado late last year and commenced post-training activities in January, at which point the mid-March release target was established.
However, the development process has not been without internal friction. Some researchers have already departed from the project, and Wang has reportedly experienced clashes with Meta veterans Chris Cox and Andrew Bosworth regarding how the new AI models should enhance the company's core advertising business operations.
Leadership Tempered Expectations Early On
Meta CEO Mark Zuckerberg had already begun managing expectations for Avocado's capabilities as early as January. During an earnings call that month, Zuckerberg acknowledged that the model would not immediately leap to the forefront of AI technology. He emphasized that while he expected it to be "good," the more critical factor would be demonstrating "the rapid trajectory we're on."
A Meta spokesperson reiterated this perspective in comments to Reuters, stating the company remains "excited for people to see what we've been cooking very soon."
Looking Beyond Avocado: The Next Model Already in Development
Despite the current delays, Meta's AI development pipeline continues to advance. According to the report, the company's next model after Avocado is already in the preliminary stages of development, internally codenamed Watermelon. This indicates that Meta is pursuing a continuous, multi-model strategy to eventually achieve parity with or surpass its rivals in the intensely competitive artificial intelligence landscape.
