Internal Leak: Alexandr Wang's First Move as Meta's New Head, Shutting Down Open-Source Large Models}

Once called the "Light of Open Source," Meta may now be heading towards a closed-source approach.

According to insiders, Meta's newly established Superintelligence Lab is currently discussing a major decision that could alter its AI development trajectory.

Within Meta, there are differing opinions on AI development paths. Some senior leaders, including new Chief AI Officer Alexandr Wang, suggest not open-sourcing their top AI models. Others believe that open-source strategies still hold advantages in the current race against competitors.

This debate centers around Meta's most powerful open-source AI model, "Behemoth" (the largest version of Llama 4). According to The Information, Meta has recently paused some work on Behemoth. Previously, delays affected two versions of Llama 4, including the inference version and the largest model (Behemoth).

Insiders reveal that Meta has completed training on Behemoth but delayed its release due to poor internal performance. Meanwhile, top executives discussed last week the possibility of abandoning this model in favor of developing a closed-source alternative.

For years, Meta has championed open-source AI models, earning widespread praise from developers. Its Chief AI Scientist Yann LeCun once said, "The winning platform will be an open platform."

However, any move to shift towards closed-source AI would represent a significant philosophical and technical change for Meta.

Meta's AI strategy shift is accompanied by major organizational changes. Last month, the company invested $14.3 billion in Scale AI, acquiring a 49% stake, and appointed Scale AI CEO Alexandr Wang as Meta’s Chief AI Officer.

Wang now leads the rebranded "Meta Superintelligence Lab," working with a core team of a dozen new researchers, several deputies from Scale AI, and former GitHub CEO Nat Friedman. The team operates in a separate office space near Zuckerberg.

Recently, Zuckerberg publicly stated in an interview that controlling more computing power is a key factor behind Meta’s successful recruitment of top talent for the Superintelligence Lab. He also announced plans to deploy a 1GW "Prometheus" supercluster in 2026, making Meta one of the first tech giants with such a large AI data center.

In response to external concerns about future directions, a Meta spokesperson stated, "Our stance on open-source AI remains unchanged. We plan to continue releasing leading open-source models. Historically, we haven't released all our developments, and we expect to continue training a mix of open and closed models."

Currently, the discussion within the Superintelligence Lab is still in early stages, and no final decision has been made. Any major change requires approval from CEO Zuckerberg. With competitors like Google, OpenAI, and Anthropic fiercely competing in AI, how Meta balances open-source principles with business interests remains a key industry focus.

During a Q&A session on Tuesday, Alexandr Wang told about 2,000 Meta AI employees that while his team’s work will remain private, the entire AI department is now committed to creating superintelligence. However, he did not specify whether the models will be open or closed source, adding uncertainty to the industry’s future.

Meanwhile, OpenAI recently announced an indefinite delay in open-sourcing its models.

This has raised concerns: if top AI companies like Meta stop open-sourcing their models, how will startups relying on open-source models for product development survive?

Some experts warn that the academic community, which heavily relies on open-source models, might also be affected. If Meta doesn’t open-source, the US academic sector could depend on Chinese open-source models instead.

However, some practitioners are already preparing for this shift by practicing distributed training approaches.

Which path will Meta choose: open-source or closed-source? The industry awaits the answer.