Mark Zuckerberg spent years telling everyone that open-source AI was the future. Now, he seems ready to change his mind. Reports indicate that Meta is working on a new model, internally nicknamed “Avocado,” that breaks with the company’s history of publicly sharing its code.
Scheduled for release in 2026, the Avocado project sits under a small team called “TBD” within Meta’s AI Superintelligence Labs. Alexandr Wang, the Chief AI Officer, leads this group. Wang reportedly prefers keeping technology under wraps, a stark contrast to the company’s previous philosophy. This shift comes amid major leadership changes, including the departure of long-time Chief AI Scientist Yann LeCun, a vocal supporter of open research who famously doubted the current direction of large language models. The company also recently cut hundreds of jobs from its Fundamental Artificial Intelligence Research (FAIR) unit.
The move away from openness follows trouble with the Llama 4 “Behemoth” model. Delays have plagued its release, and executives reportedly considered scrapping it entirely after developers showed little enthusiasm for early versions. Zuckerberg himself warned earlier this year that while Meta would lead in open source, it wouldn’t give away everything, especially when it comes to safety and superintelligence.
This represents a massive U-turn for Zuckerberg. Just last year, he wrote a manifesto declaring open source as the “path forward” and dismissing closed platforms with colorful language. But the pressure to compete with OpenAI and Google is intense. With plans to pour a reported $600 billion into AI development over the coming years, the CEO appears willing to lock down his technology if it ensures Meta doesn’t fall behind. The days of Meta freely handing out its most powerful tools may be coming to an end.
