Mark Zuckerberg has spent the past year telling investors and developers that Meta Platforms intends to push the boundaries of artificial intelligence. But inside the company, progress on its next major AI model has reportedly been slower than expected.

People familiar with the work say Meta’s upcoming foundational AI model — internally called Avocado — is not yet performing at the same level as some of the most advanced systems from competitors.

During internal testing, the model reportedly struggled to match the capabilities of models developed by companies such as Google, OpenAI and Anthropic, particularly in tasks involving reasoning, coding and writing.

That does not mean the project has stalled. Sources say Avocado shows clear improvements compared with Meta’s previous AI systems. It reportedly performed better than Gemini 2.5, an earlier model from Google released earlier this year. However, it still falls behind the more advanced Gemini 3.0, which set a higher benchmark when it arrived later.

Because of those gaps, Meta has pushed back the release timeline. The company had originally hoped to launch the model this month, but the rollout is now expected no earlier than May.

The delay has also triggered internal discussions about whether Meta should temporarily license Google’s Gemini technology to support some of its AI services. According to people familiar with the conversations, no final decision has been made.

The performance of Meta’s next AI model matters more than ever. Foundational models serve as the backbone for many modern AI tools, including chatbots, coding assistants and video generators. Companies that lead in this space often attract the best engineers and move faster in developing new products.

For Zuckerberg, the stakes are particularly high. The 41-year-old CEO has made artificial intelligence central to Meta’s long-term strategy. The company, which operates platforms like Facebook, Instagram and Threads, has invested heavily in AI research, hiring top scientists and expanding data-center infrastructure to support the technology.

Meta’s spending plans reflect that push. Earlier this year, the company said it could spend up to $135 billion in 2026, nearly doubling last year’s total.

Industry experts note that delays in AI development are common. Building and refining large models can take longer than expected, especially as companies aim to match or surpass rivals.

Zuckerberg himself has hinted at a longer timeline in recent months. Speaking to investors in January, he suggested that Meta’s next model may be more important as a step forward rather than an immediate leap ahead of competitors.

Meta doubled down on its AI ambitions after its previous system, Llama 4, failed to meet expectations. In response, the company invested $14.3 billion in the startup Scale AI and brought its CEO, Alexandr Wang, into Meta as chief AI officer.

Wang helped establish a new internal research group known as TBD Lab, short for “to be determined.” The team began building two projects with fruit-themed names: Avocado, a large language model, and Mango, a system focused on generating images and video.

Avocado completed its initial pretraining phase late last year. The team then moved to post-training earlier this year, a stage that focuses on refining the model’s responses and preparing it for release.

So far, the division has released only one product — Vibes, an AI video app similar to Sora from OpenAI.

Another debate inside the company centers on whether the next model should be open source or closed. Meta has traditionally supported open AI models that developers can build upon, while some competitors argue that restricting access helps reduce safety risks.

Sources say Zuckerberg and Wang have recently leaned toward keeping the new model closed.

The AI lab itself is still evolving. The group currently has around 100 employees and continues to recruit talent, though a few researchers have left before the Avocado launch.

There have also been disagreements among executives about how the new AI technology should support Meta’s advertising business. Those discussions involve senior leaders including chief product officer Chris Cox and chief technology officer Andrew Bosworth.

Last week, Meta announced plans to create a dedicated AI engineering team under Bosworth that will work closely with Wang’s division.

Rumors recently circulated online suggesting tensions between Zuckerberg and Wang. Meta quickly denied those claims. In response, Zuckerberg shared a photo of the two together at the company’s headquarters on Threads, appearing to dismiss the speculation.

Even as work continues on Avocado, Meta is already thinking about what comes next. According to people familiar with the plans, the company’s future AI model may follow the same naming pattern — with an even bigger fruit-inspired name: Watermelon.

Also Read: Meta Launches Applied AI Engineering Team to Advance Superintelligence Efforts