Meta Pivots from Llama to Closed AI Models, Abandoning Open Source Roots

Abandoning the open-source philosophy that defined its artificial intelligence strategy, Meta is reportedly pivoting to a closed, proprietary model codenamed “Avocado.” The shift follows the internal failure of its Llama 4 “Behemoth” model and marks a capitulation to competitive pressure.

Driving this reversal is a newly installed leadership team led by Chief AI Officer Alexandr Wang. The former Scale AI CEO, now heading the elite “TBD Lab,” has discarded the company’s academic roots for a secretive culture that has sparked significant internal attrition.

Fueling this reset is a massive capital injection, with the company raising its 2025 capital expenditure (Capex) guidance to $72 billion to build the infrastructure required for its new closed garden.

The ‘Avocado’ Pivot: Closing the Garden

Meta is fundamentally altering its AI strategy, moving from the open-source Llama series to a proprietary model codenamed “Avocado,” targeting a Q1 2026 release.

This strategic reversal directly contradicts CEO Mark Zuckerberg’s previous public stance, where he argued that open source was “closing the gap” with closed models. In his open source manifesto, Zuckerberg had championed the approach as the industry’s future.

However, the internal failure of the Llama 4 launch appears to have forced a reassessment. Postponed in May 2025, the flagship “Behemoth” model was shelved after underperforming on critical benchmarks, leaving the company without a competitive answer to rivals.

Precipitating this strategic reversal was the rapid rise of Chinese competitor DeepSeek. Leadership was reportedly spooked after DeepSeek’s R1 model successfully copied Llama’s architecture, highlighting the commercial risks of releasing open weights that can be easily cloned by adversaries.

Zuckerberg has now adopted a more cautious tone regarding future releases. Speaking on the company’s strategy, he noted already in July that Meta needed “to be rigorous about mitigating these risks and careful about what we choose to open source.”

Leaving the future of the Llama brand in limbo, the decision creates uncertainty for developers. It remains unclear whether the name will continue as a “lite” offering or be deprecated entirely in favor of the new proprietary line.

The ‘TBD Lab’ Takeover

A new power structure has emerged within Meta’s AI division, led by Chief AI Officer Alexandr Wang and Head of Product Nat Friedman. The pair now control the company’s AI destiny, sidelining long-time executives. Most notably, Chief Product Officer Chris Cox was removed from AI oversight following the Llama 4 debacle.

Under their leadership, a “demo, don’t memo” culture has taken hold within the secretive Superintelligence TBD Lab. This approach bypasses the rigorous peer-review processes favored by the Fundamental AI Research (FAIR) team, prioritizing speed over academic validation.

Wang defends the streamlined approach as necessary for agility. In an internal memo regarding restructuring, he argued that “by reducing the size of our team, fewer conversations will be required to make a decision, and each person will be more load-bearing and have more scope and impact.”

Compounding the internal friction is a conflict between the new guard and veteran researchers. The cultural clash has led to the departure of Chief AI Scientist Yann LeCun, whose departure signals a definitive break from the company’s research-driven past.

Friction is palpable among the rank and file. Former researcher Tijmen Blankevoort described the internal environment in direct terms, stating that “it’s not just dysfunction – it’s a metastatic cancer that is affecting the entire organisation.”

Operating almost as a separate startup, the “TBD Lab” has isolated itself from the broader organization. Its members reportedly do not even use Meta’s internal Workplace”communication tools, further deepening the divide between the elite unit and the rest of the company.

Buying a Way Out

Meta has been trying to spend its way out of its development crisis so far, raising its 2025 capital expenditure guidance to a dramatic $70–$72 billion.

A significant portion of this capital is funding the “Prometheus” project, a gigawatt-scale data center in Ohio featuring on-site power generation to support large-scale training runs.

Underpinning this significant pivot is a willingness to pay any price for infrastructure. The company’s desperation was evident in the $14.3 billion investment for a 49% stake in Scale AI, a deal that shattered the data labeling firm’s neutrality.

Human Toll: A ‘Metastatic’ Culture

Redefining the cost of acquisition in the AI sector, the “buy or poach” strategy has become Meta’s primary recruiting tool this year. The company has been offering compensation packages reportedly reaching nine figures to lure top researchers away from competitors.

Internally, the disparity between the new “dream team” hires and existing staff has reportedly created a two-tier system.

Contrasting sharply with the allegations of a ‘culture of fear’ among long-time employees, the lavish bonuses for external recruits have fueled resentment.

Layoffs of 600 employees from the legacy FAIR and infrastructure teams in October 2025 further deepened the divide. Signaling a willingness to cut deep into institutional memory, the recent layoffs demonstrated that the company will sacrifice its past to fund its new direction.

Retention rates at the company have plummeted below those of its rivals. While money can buy talent, the ongoing exodus suggests it cannot necessarily buy loyalty or culture.