**The Delays and Benchmark Controversy** Last year, Meta released Llama 4 in April 2025, which immediately sparked a benchmark manipulation controversy. Its "Maverick" version (using 17 billion active parameters out of 400 billion total) did not impress the market. At the same time, Meta had already finished pre-training a massive 288 billion parameter model called "Behemoth" in 2025, which served as a teacher model for Llama 4.

The scandal finally came to a head in January 2026, when Chief AI Scientist Yann LeCun publicly admitted to the manipulation. They didn't just use different versions on different tests—Meta used a powerful, unreleased version of Llama 4 for their biggest marketing test (LM Arena) to get high scores, while shipping a noticeably weaker version to actual users. That specific detail is what made the AI community furious.

Because of this tough competition and controversy, Meta realized they needed a much stronger next model. Their next big project, secretly codenamed 'Avocado', finished basic training in early 2026, but the public launch has been pushed back to at least May 2026.

**Massive Spending and Infrastructure Changes** According to reports from Reuters and The New York Times covering their recent earnings, Meta is drastically upgrading its AI infrastructure with a total budget surging to $115–$135 billion.

In February 2026, as reported by Reuters, they signed a massive official deal worth up to $60 billion for AMD AI chips over 5 years. This is part of a larger $100 billion+ infrastructure setup. They are also renting large Google TPU clusters for extra computing power.

Leadership has completely changed as well. After investing $14.3 billion in Scale AI for a 49% non-voting stake—a move confirmed by outlets like CNBC and Reuters—its founder Alexandr Wang joined Meta as the new Chief AI Officer. Meanwhile, veteran AI scientist Yann LeCun left the company following 600 job cuts at the newly formed Meta Superintelligence Labs in October 2025. Despite this internal chaos and recent delay news, Meta's public statements officially claim training is "on plan, with no meaningful changes."

**Moving Away from Fully Open AI** When Avocado finally launches—whether it officially carries the 'Llama 5' name or not—it is expected to be dramatically more efficient. Internal Meta documents, widely reported, suggest Avocado is: * 10 times more compute-efficient than Llama 4 Maverick for text tasks. * Over 100 times more efficient than the massive Behemoth model.

The biggest shift is Meta's approach to open-source. In the past, Meta released "open-weights" models, meaning the core AI files were publicly available to download. Now, according to *The Information*, Meta will lock up their absolute largest AI, keeping it closed and likely offering it as a paid service like their competitors.