Meta's AI Ambitions Hit Another Wall

Mark Zuckerberg announced early this year that Meta would spend between $115 billion and $135 billion in pursuit of AI superintelligence. The results, according to leaked internal evidence, are not yet matching the ambition.

Meta's next-generation text AI model, codenamed Avocado, has been pushed back again — this time to at least May 2026. The delay comes after internal testing revealed the model still falls short of frontier systems from Google, OpenAI, and Anthropic. More surprisingly: Meta appears to be quietly routing some user requests through Google's Gemini models in the meantime.

What the Internal Evidence Shows

According to TestingCatalog, which gained access to the Meta AI interface's internal model selector, the company is currently running parallel experiments across multiple Avocado variants simultaneously. The configurations under evaluation include:

• Avocado 9B — a smaller 9 billion parameter variant
• Avocado Mango — a multimodal variant with "agent" and "sub-agent" labels, capable of image generation
• Avocado TOMM — described as a "Tool of Many Models" built on top of Avocado
• Avocado Thinking 5.6 — the latest version of Meta's reasoning-focused model variant
• Paricado — a text-only conversational model

The proliferation of variants suggests Meta hasn't settled on a final configuration — a significant development given the model was originally targeted for Q1 2026.

The Gemini Fallback

Perhaps the most striking revelation: some requests within the Meta AI interface are already being routed through Google's Gemini models. The report says Meta's AI leadership has "reportedly discussed temporarily licensing Google's Gemini technology," and that the current A/B testing of Gemini-based responses alongside Avocado versions "confirms that this is not merely theoretical."

This puts Meta in an awkward position — the company that spent years building open-source Llama models as an alternative to closed commercial AI may be quietly depending on a competitor's system to keep its consumer products competitive.

A Capability Gap That Can't Be Ignored

The report notes that Avocado can solve complex math problems that earlier Llama models couldn't handle — but that these same problems had already been solved by Gemini 3 and GPT months earlier. That's the core problem: Meta is running hard but not catching up fast enough.

Meta has also reportedly reversed course on its open-source philosophy. Avocado is expected to be proprietary — a stark departure from the Llama era that made Meta a darling of the open-source AI community.

"As we've said publicly, our next model will be good but, more importantly, show the rapid trajectory we're on." — Meta spokesman Dave Arnold

Stakes for Meta's 3 Billion Users

Meta AI reaches users across Facebook, Instagram, WhatsApp, and its Ray-Ban smart glasses — a massive distribution advantage that no purely AI-focused company can match. Even a somewhat-behind Avocado would represent a meaningful upgrade for the hundreds of millions of users interacting with today's Llama-based responses.

But the combination of delays, reliance on a competitor's model, and a pivot away from open-source raises uncomfortable questions about whether Meta's bet on AI superintelligence is delivering results fast enough to justify the investment — or whether Zuckerberg's push for dominance is outpacing the company's technical execution.