Meta’s AI Strategy is Bigger Than Just Llama

In the fast-developing frontier LLM race, Meta has been playing relatively silently. 

In April last year, it released the Llama 4 family of open-source models—Scout and Maverick—best suited for multimodal tasks and context-heavy applications like summarising books, long reports, and large datasets. While it has delayed the release of the 2 trillion parameter Behemoth model, as Meta engineers reportedly weren’t able to significantly improve its capabilities over the existing models, its competitors in the AI ecosystem have been dishing out mixture-of-experts language models every month.

OpenAI released o3-pro, positioning it as the highest-performance reasoning model in the o3 family. Google released Gemini 3.1 Pro in February this year, best suited for visual reasoning and analysing long-form videos. Meanwhile, DeepSeek-V3 is an affordable alternative for AI-powered coding and debugging, and API deployment in enterprise applications. The Hangzhou, China-based company is also expected to release the V4 model this month.  

But Meta’s silence may now be coming to an end. At the India AI Impact Summit, Meta’s Chief AI Officer Alexandr Wang said the Big Tech company was preparing to roll out a new series of AI models this year, with the first expected in the next couple of months. These models, he said, will be deeply integrated across Meta’s products.

According to reports, the two new foundation models are codenamed Mango and Avocado. 

Mango is being developed as a multimodal system that can generate and understand images and videos, and is expected to power visual features across Meta’s apps and services. Avocado, meanwhile, will serve as the company’s new large language model, with improvements in text generation, reasoning, and coding compared to its current systems. 

For Wang, however, the modes are just one piece of a much larger puzzle.

“Our vision is personal superintelligence—AI that knows you, your goals, your interests, and helps you with whatever you’re focused on doing,” Wang said.

Unlike generic assistants, Wang envisions AI systems deeply integrated into users’ lives, helping manage health goals, organise events, coordinate travel, and even strengthen relationships. “It won’t just do your admin. It’ll be an extension of you so you can be you more,” he said.

Separately, Meta is testing a new AI shopping assistant in the Meta AI app to help users find products and compare different options. 

In a conversation with AIM, Kiumarse Zamanian, VP Analyst at Gartner, says Meta is trying to balance profits versus building models for the public good. “Meta’s long-term LLM strategy appears to use selective openness as a tool for scale and influence, while building the commercial, technical, and policy foundations for enterprise and government adoption,” he notes. 

He adds that enterprises should expect a hybrid model as Meta continues releasing open-weight models such as Llama 4 while investing in proprietary infrastructure and partnerships.

AI Leadership 

Over the past year, Meta has reshaped its AI leadership and strategy. 

Wang has taken charge of the Superintelligence team, steering the company’s broader AI efforts. He was acqui-hired by the company following a $14.3 billion investment last year in data-labelling unicorn Scale AI.

Meanwhile, longtime Chief AI Scientist Yann LeCun has moved on to start a new venture centred on world models, an approach designed to help AI systems better understand and interact with the real world.

The Mark Zuckerberg-led company has also been shifting its AI priorities. In October last year, it laid off roughly 600 people within its AI divisions, including the Fundamental AI Research (FAIR) unit, which focuses on long-term research.

Julien Chaumond, CTO of Hugging Face, expressed uncertainty about Meta’s AI direction in a recent LinkedIn post. “I really, really wonder what Meta AI is up to these days,” he wrote, adding, “I feel they might either ship the mother of all AI bangers or fade themselves into oblivion.”

Meta’s Compute Efforts 

Even though Meta didn’t release any new models, it has been steadily improving its compute capabilities. Earlier this year, the company launched a new top-level initiative called Meta Compute to build a large-scale computing infrastructure for its long-term AI plans.

The company also plans to invest up to $600 billion in US AI data centres through 2028, with 2026 capex potentially touching $135 billion. It has already broken ground on a 1 GW campus in Indiana, US, alongside massive projects such as the 5 GW Hyperion facility in Louisiana and the 1 GW Prometheus supercluster in Ohio.

Meta has also firmed up major multi-year partnerships across the chip ecosystem. 

It expanded its relationship with NVIDIA to deploy CPUs, millions of Blackwell and Rubin GPUs, and Spectrum-X networking gear across new hyperscale data centres built for AI training and inference. It also signed a deal with AMD for up to 6 GW of Instinct GPUs, starting with custom MI450 chips, along with EPYC processors, with deliveries beginning in the second half of 2026. 

In parallel, The Information reported that Meta has entered a multi-billion-dollar agreement with Google to rent its tensor processing units for next-generation AI model training.

“Meta’s AI chip and data centre investments will become a true competitive advantage if they are integrated with workload-aware system design, resilient supply chains, energy-efficient infrastructure, platform operations, organisational transformation, and ecosystem partnerships,” says Zamanian.

He adds that taking this broader approach could help Meta compete more strongly, not only in AI research but also in building AI products that are scalable, reliable, and profitable.

On the data front, Meta has struck a multi-year AI licensing partnership with News Corp that could be worth up to $50 million annually. The deal allows Meta to use content from News Corp titles such as The Wall Street Journal and New York Post to train and improve its models. 

New Wearables 

Meta’s wearable strategy is gaining commercial momentum. 

Ray-Ban maker EssilorLuxottica more than tripled sales of its Meta artificial intelligence glasses last year. The French-Italian eyewear group reported selling more than 7 million AI glasses in 2025, compared with a combined 2 million units sold across 2023 and 2024, according to its previous quarterly disclosure.

“Meta’s wearables are a core part of the company’s AI roadmap. They serve as ‘edge substrates’ for AI agents, enabling hands-free, contextual, multimodal interactions and collecting valuable data to fuel AI development,” says Zamanian.

According to recent reports, Meta is reviving its smartwatch project (codenamed Malibu 2) for a late-2026 launch. The device is expected to include health tracking and a built-in Meta AI assistant. 

Together, these moves show that Meta is serious about building an AI hardware ecosystem, which could give the company more direct access to users.