Three months into 2026, Meta’s AI ambitions are unraveling. The company’s flagship AI model “Avocado” has been pushed to May after failing to match competitors. Leadership is in open conflict. And in the most startling sign of desperation, executives have discussed licensing Google’s Gemini to power Meta’s own products.
For a company that spent years positioning open-source Llama models as the alternative to proprietary AI, the reversal is striking.
The Avocado Problem
Meta’s “Avocado” model was supposed to launch in March 2026 as the company’s most capable AI system. Instead, internal testing revealed it trails Google, OpenAI, and Anthropic in logical reasoning, programming, and writing tasks.
According to Trending Topics, the model performs better than Meta’s previous systems and beats Google’s older Gemini 2.5. But it cannot compete with Gemini 3.0, which launched in November 2025.
The delay puts Meta’s $115-135 billion AI spending commitment under scrutiny. As PYMNTS reports, unlike Amazon, Microsoft, and Google, Meta lacks a cloud business to directly monetize its AI infrastructure. The company depends on improving ad targeting, content recommendations, and assistant features across Facebook, Instagram, and WhatsApp.
The Gemini Consideration
The most jarring detail from the Avocado delay: Meta’s AI leadership has discussed temporarily licensing Google’s Gemini to support its products while Avocado improves.
No decision has been finalized. But the mere discussion inverts years of strategy. Meta spent heavily positioning Llama as the open-source alternative to proprietary systems from OpenAI and Google. CEO Mark Zuckerberg publicly argued that open source was “closing the gap” with closed models.
If Meta licenses Gemini, users of Meta AI across WhatsApp, Instagram, and Facebook would receive responses generated by Google’s infrastructure, branded as Meta AI, with no visible indication of the switch.
LeCun’s Exit and Public Criticism
The strategic chaos coincides with the departure of Yann LeCun, one of AI’s most recognized figures and Meta’s Chief AI Scientist since 2013.
LeCun left in November 2025 to found AMI Labs in Paris, which raised $1.03 billion at a $3.5 billion valuation in March 2026. His new venture focuses on “world models” - systems that understand physical reality rather than just processing text.
Since leaving, LeCun has been publicly critical of his replacement. In a Financial Times interview, he called Alexandr Wang, Meta’s new Chief AI Officer, “inexperienced” and predicted “a lot of people” would leave the company.
“There’s no experience with research or how you practice research,” LeCun said, according to Semafor. “You don’t tell a researcher what to do. You certainly don’t tell a researcher like me what to do.”
LeCun also revealed that Meta’s AI team “fudged” some benchmark results for Llama 4, which angered Zuckerberg and caused him to lose confidence in the team.
The Leadership Shakeup
Wang joined Meta in June 2025 after Zuckerberg’s $14.3 billion investment for a 49% stake in Scale AI. At 28, the Scale AI co-founder now co-leads Meta Superintelligence Labs alongside Nat Friedman, former CEO of GitHub.
The appointment restructured Meta’s entire AI organization. Hundreds of FAIR (Facebook AI Research) employees were laid off. The focus shifted from open research to proprietary model development.
The tension between Wang and the research community LeCun represented reflects a deeper strategic question: Is Meta’s AI problem a research problem or an execution problem?
LeCun’s position is clear. He maintains that simply scaling up large language models will not lead to superintelligence - a view that puts him at odds with Meta’s current direction. His departure, and the $1 billion backing for his alternative approach, suggests the research community has doubts about the LLM-scaling path Meta is pursuing.
The Open Source Reversal
Meta’s move away from open source began in late 2025 after DeepSeek’s R1 model demonstrated how easily Llama’s architecture could be copied. Leadership was reportedly “spooked” by the commercial risks of releasing open weights that adversaries could clone.
The new strategy centers on proprietary models:
- Avocado: Text and code generation (delayed to May 2026)
- Mango: High-resolution image and video generation
- Watermelon: Planned successor to Avocado
Llama 4 “Behemoth,” originally expected in early 2025, remains unreleased after multiple delays. Internal sources reported “serious concerns” about whether the model delivered enough improvement over earlier versions.
What This Means
Meta’s AI crisis reflects broader industry dynamics. Building frontier AI models has proven more difficult than scaling infrastructure. Google, OpenAI, and Anthropic have maintained leads despite Meta’s unprecedented spending.
The company faces an uncomfortable reality: its AI capabilities do not match its AI expenditure. Licensing from Google, even temporarily, would be an acknowledgment that building competitive AI in-house is harder than expected.
For the broader AI ecosystem, Meta’s struggles raise questions about the future of open-source AI. If the most well-funded open-source effort cannot compete with proprietary models, alternatives like Alibaba’s Qwen, Mistral, and smaller open-weight projects face an uphill battle.
The Bottom Line
Meta is spending more on AI than any company except Microsoft, yet it is falling behind. The company’s response - abandoning open source, reorganizing leadership, and potentially licensing from a competitor - suggests fundamental strategic uncertainty. Whether Wang and Friedman can execute where LeCun’s research organization could not remains the $135 billion question.