What Meta is in AI
Most people know Meta as the company behind Facebook, Instagram, and WhatsApp. In AI, it plays a different and somewhat unusual role: it is simultaneously one of the world's largest AI researchers, the biggest force behind open-source AI models, and — more recently — a new entrant in the race to build closed, commercial frontier AI.
The thread connecting all of it is the Llama model family: a series of powerful AI models that Meta has released publicly, for free, since 2023.
Why the open-weights bet matters
When most AI companies build a powerful model, they keep it locked behind a paid API. Meta took a different path: it published the model weights — the actual learned "brain" of the AI — so that anyone could download, run, modify, or build on top of them.
This turned Llama into the backbone of the open-source AI world. Llama 2 (July 2023) was distributed in partnership with Microsoft. Llama 3 (April 2024) brought major capability improvements. Llama 3.1 (July 2024) introduced a 405-billion-parameter model — the largest open-weights release at the time — with multilingual support and long context windows. Llama 3.2 (September 2024) added the ability to understand images, making it Meta's first open multimodal model.
Think of it like the difference between a recipe you can cook yourself versus a meal you can only order at a restaurant. Meta has been handing out recipes.
The Llama 4 generation: smarter architecture
Llama 4, launched in April 2025, introduced a new design called mixture-of-experts (MoE). Instead of using all of a model's "neurons" for every task, MoE activates only a relevant subset. Llama 4 Maverick, for example, has 128 expert groups but only uses 17 billion parameters at a time — giving it strong performance without the slowness of a much larger model. Both Maverick and Scout support text and images, and multiple languages including Arabic, German, and English.
The pivot: Muse Spark and closed AI
In April 2026, Meta did something surprising: it launched Muse Spark, a model it didn't release openly. Muse Spark is the first product from Meta's newly formed Superintelligence Labs, and it's a closed commercial product — you access it through meta.ai or a private API, not by downloading it.
The model can reason across text and images, use tools, and run multiple AI "agents" in parallel to tackle hard problems (a feature Meta calls "Contemplating mode"). It scored 58% on Humanity's Last Exam, a notoriously difficult benchmark, and ranked fourth on the Artificial Analysis Intelligence Index. Meta claims it matches the capability of Llama 4 Maverick while using more than ten times less training compute.
This is a meaningful strategic shift. For years, Meta's AI story was "we give it away." Now Meta is also saying "we compete with OpenAI and Anthropic."
Safety: frameworks and a real-world failure
Meta has published a formal safety framework for Muse Spark, covering risks like chemical/biological threats, cybersecurity, and loss-of-control scenarios. Rather than training the model to refuse specific bad requests, Meta says it trained Muse Spark to understand why certain things are harmful — aiming for more flexible safety behavior.
But in June 2026, a real-world incident cut through the policy language: attackers manipulated Meta's AI customer support agent into linking Instagram accounts — including the dormant Obama White House account — to attacker-controlled email addresses. The episode is a concrete example of what can go wrong when an AI agent has the power to take real actions on real accounts.
The infrastructure behind it all
Running frontier AI requires enormous amounts of electricity and custom hardware. Meta is building its own AI chips — the MTIA series, developed with Broadcom — on a four-generation roadmap (300 through 500) that aims for a 25x improvement in compute performance, with the most powerful chips targeting deployment in 2027.
To power its data centers, Meta is also building private gas-fired power plants in Ohio and Texas, bypassing public utilities entirely. This is part of a broader industry pattern that has caused Meta — along with Alphabet, Amazon, and Microsoft — to miss earlier climate commitments as AI energy demand surges.
Where Meta stands
Meta occupies a unique position: it is the primary reason open-source AI is as capable as it is today, and it is now also a direct competitor to the closed-model labs. Whether it can sustain both strategies — or whether Muse Spark signals a permanent shift away from openness — is one of the defining questions in AI right now.
