After spending billions and enduring significant staff turnover to revamp its AI unit, Meta is weighing a move to make its new model, nicknamed Avocado, closed source, multiple outlets reported.

Since open-source Llama 4 received mixed reviews following its April release, Meta CEO Mark Zuckerberg has embarked on an ongoing crusade to reverse the company’s AI fortunes. Meta aqui-hired Scale CEO, Alexandr Wang, and co. went on an AI hiring spree to build, in Zuckerberg’s telling, “the most elite and talent-dense team in the industry.” Wang’s team, one of a few AI teams within Meta, is known internally as TBD Lab. While pay packages reportedly stretching up to nine figures have attracted talent, the group has reportedly not cohered.

Meta’s AI restructuring led some researchers to depart for other AI labs. Meta’s chief AI scientist, Yann LeCun, left in November to found his own startup. CNBC has reported that LeCun was rankled by the 600 layoffs that Meta Superintelligence Labs (MSL) underwent in October. Just this week, MSL employees Sang Michael Xie and Vitaliy Chiley decamped from the company.

Those still at Meta have reportedly clashed on direction. Senior executives said some of Meta’s AI efforts should be oriented toward improving the company’s social media and advertising businesses, while Wang argued Meta should catch up with frontier models offered by OpenAI and Google before focusing on products, per The New York Times. The Times also reported that budget cuts to Meta’s metaverse team were rerouted to Wang’s unit.

TBD Lab’s new frontier model, being closed source, would represent a major strategic shift for Meta, which has long touted its open-source AI efforts. The risks of open-source development became clear when DeepSeek’s R1 model incorporated components of Llama’s architecture, which angered some at Meta, CNBC reported.

Open-source cuts both ways, though. TBD Lab is using Alibaba’s Qwen model as part of the training process for Avocado, Bloomberg reported.

Our Deeper View

Meta has fallen behind other open-source AI models, so it’s pivoting to the even more competitive field of closed source AI models. Keeping Llama open-source could have created a wedge for Meta to differentiate itself from other hyperscalers, but as things currently stand, it’ll be up to Wang to lead Meta’s proprietary model past those offered by OpenAI, Google, and Anthropic. In the words of Wang’s unit within Meta, the odds of that happening are very much TBD.