After “Avocado”, Meta is now betting on “Mango” in the AI ​​race, according to the newspaper

• Under pressure from the White House for his Chinese origin: TikTok says it has signed agreements to create a new company in the United States
• Bet: Is China’s accelerated progress in robotics forming a bubble?

According to the WSJ, citing people familiar with the matter, the social media giant is working on a high-impact AI model for images and videos called “Mango.” At the same time, the company is developing “Avocado”, its next major text language model (LLM).

The new roadmap was reportedly detailed during an internal Q&A session held today, with participation from Meta’s Chief AI Officer, Alexandr Wang, and Chief Product Officer, Chris Cox. The models are expected to go on sale in the first half of 2026, signaling a long-term strategy to move beyond incremental updates to the current Llama family.

• Pass negotiated as a millionaire athlete: AI talents receive proposal of up to 1 billion reais

The development of Mango and Avocado represents the first major result of Meta Superintelligence Labs (MSL), a specialist division created after a major restructuring last year.

To lead the group, CEO Mark Zuckerberg personally orchestrated high-profile recruiting efforts, attracting more than 20 researchers from OpenAI and hiring Wang, the 28-year-old founder of Scale AI, to oversee the initiative.

The initiative highlights Zuckerberg’s desire to spend aggressively to narrow the gap with his rivals. As part of the transition, Meta reportedly invested more than $14 billion to acquire a near-controlling stake in Scale AI, cementing Wang’s role as chief architect of Meta’s strategy in the post-Llama period.

• “Us versus them”: Meta’s new AI superstars begin to clash with the rest of the company

During the internal session, Wang emphasized that Avocado will place a strong emphasis on advanced programming capabilities, a traditionally weak point of Meta’s previous models compared to its competitors.

Even more ambitiously, Wang noted that Meta is in the early stages of developing “global models.” Unlike current LLMs, which predict the next word in a sequence, world models seek to understand physical reality by processing enormous volumes of visual information.