Meta Reportedly Delays Release of New AI Model Avocado: Performance Lags Behind Gemini 3

The New York Times reports that Meta’s latest AI model, Avocado, has experienced development delays. Its performance in internal testing has not yet caught up with Google’s latest Gemini 3, leading to a postponement of the original March release schedule to at least May. Meanwhile, Meta is also discussing the possibility of temporarily adopting Google’s Gemini model as a transitional solution to maintain its AI product competitiveness.

AI Race Turns Uncertain: Meta’s New Model Avocado Delayed Until After May

According to the New York Times, Meta has been actively developing a new generation of foundational AI models called “Avocado,” which was originally scheduled for release in March this year. However, the launch has now been pushed back to at least May, possibly even June.

Sources indicate that in internal tests focusing on reasoning ability, programming, and text generation, Avocado has not yet reached the same level as flagship AI models from competitors like Google, OpenAI, and Anthropic. As a result, the company has decided to delay the release to allow more time for optimization.

Performance Below Expectations: Ranges Between Gemini 2.5 and Gemini 3

Additionally, Reuters cited sources revealing that Avocado’s overall performance is indeed better than Meta’s previous AI models and has surpassed Google’s Gemini 2.5 in some tests. However, it still falls short of the Gemini 3 released at the end of last year.

This suggests that in the current AI foundational model competition, Google, OpenAI, and Anthropic remain leaders in areas such as video generation, chatbots, and programming. For Meta, AI model performance not only affects product functionality but also impacts its ability to attract top research talent and establish technological leadership.

Meta Considers Using Gemini as a Transitional Solution

The report also notes that Meta’s AI division has discussed temporarily licensing Google’s Gemini model to support some of its AI products before Avocado is ready. However, no final decision has been made yet. This move highlights that even large tech companies may adopt collaboration or licensing strategies to stay competitive in the fierce AI race.

Previously, Google and Meta collaborated on a project called “TorchTPU,” aimed at enabling TPU chips to natively and efficiently support the popular AI development framework PyTorch, challenging NVIDIA’s long-standing software moat built around CUDA.

(Google and Meta Join Forces to Challenge NVIDIA’s Monopoly with TorchTPU—Can It Empower TPUs to Weaken GPUs?)

Meta Invests Heavily in AI, Aiming for “Superintelligence”

Meta CEO Mark Zuckerberg has been increasing investments in artificial intelligence in recent years, viewing it as a core part of the company’s future development. The company has poured substantial funds into building AI data centers and infrastructure, with capital expenditure estimates for this year reaching between $115 billion and $135 billion.

Meta has also recruited Alexandr Wang, founder of Scale AI, as head of AI operations, and established a new research team to develop cutting-edge AI models, including Avocado. This aligns with its “superintelligence” vision—aiming for AI capabilities to surpass human intelligence.

This article on Meta delaying the launch of its new AI model Avocado, with performance lagging behind Gemini 3, first appeared on Chain News ABMedia.