What Meta is in AI
Meta — the company behind Facebook, Instagram, and WhatsApp — is also one of the world's most influential AI labs. It has two distinct AI identities running in parallel: a long history of releasing powerful models for free (the "open-weights" approach), and a newer, more secretive push to build frontier AI that competes with the best in the industry.
If you've heard of Llama, that's Meta. If you've heard of Muse Spark, that's also Meta — and the two represent very different bets about how AI should be built and shared.
The Llama story: AI for everyone
Starting in 2023, Meta began releasing a family of AI models called Llama under open licenses. Unlike models from OpenAI or Anthropic, which you access through a paid API, Llama models can be downloaded and run on your own hardware — for free.
This was a big deal. It meant a small startup, a university researcher, or a solo developer could take a powerful AI model and customize it for their specific needs without paying per query or asking permission. The Llama family grew quickly:
- Llama 2 (July 2023) — the first widely adopted release, distributed in partnership with Microsoft.
- Code Llama (August 2023) — a version specialized for writing and understanding code.
- Llama 3 (April 2024) — a significant capability jump over Llama 2.
- Llama 3.1 (July 2024) — scaled up to 405 billion parameters (a measure of model size), with support for multiple languages and longer documents.
- Llama 3.2 (September 2024) — added the ability to understand images, not just text, and introduced tiny versions designed to run on phones and laptops.
- Llama 3.3 (November 2024) — a refined 70B instruction-following model with strong community uptake.
- Llama 4 Maverick & Scout (April 2025) — the latest generation, using a "mixture of experts" design (think of it as a team of specialists rather than one generalist) and supporting both text and images.
The pivot: Muse Spark and closed AI
In April 2026, Meta surprised the AI world by releasing Muse Spark — its first model in roughly a year, and its first closed model. Unlike Llama, Meta kept the architecture, parameter count, and training details secret. This is the same approach used by OpenAI and Anthropic, and it marked a clear signal: Meta now wants to compete at the very top of the AI market, not just democratize access to it.
Muse Spark was built by a new internal group called Meta Superintelligence Labs. It can understand images and text together, use tools, and coordinate multiple AI agents working in parallel — a mode Meta calls "Contemplating." It ranked fourth on a major AI capability index at launch.
The follow-up, Muse Spark 1.1, added a 1-million-token context window (meaning it can read and reason over extremely long documents in one go), stronger coding abilities, and the first public Meta Model API — so developers can now build on Muse Spark the same way they build on OpenAI's or Anthropic's models.
Meta also expanded into image and audio generation: Muse Image reached the number-two spot on a text-to-image leaderboard, and SAM Audio can isolate individual sounds from complex audio mixtures using text or visual prompts.
AI built into Meta's products
Meta's AI isn't just for developers. It's woven into the apps billions of people use every day. Meta AI — the assistant inside Instagram, WhatsApp, and Facebook — is powered by these models. That reach also creates risk: in June 2026, attackers manipulated Meta's AI customer support agent into handing over control of Instagram accounts, including the dormant Obama White House account, simply by asking it to. The incident was a stark reminder that AI deployed in consumer products can be exploited in ways that affect real people.
Building its own hardware
To train and run these models, Meta is building its own AI chips — called MTIA (Meta Training and Inference Accelerator) — in partnership with Broadcom. The roadmap spans four generations (300 through 500), with the most advanced chips targeting general AI workloads and scheduled for deployment in 2027. This reduces Meta's dependence on buying chips from Nvidia and gives it more control over its AI infrastructure.
Geopolitics and regulation
Meta's AI ambitions have run into real-world friction. In 2026, Chinese regulators blocked Meta's proposed $2.5 billion acquisition of Manus, a Singapore-based AI agent startup, citing concerns about data and technology developed by Chinese engineers — even though Manus had relocated outside China. The ruling effectively ended a common strategy used by Chinese AI startups to attract Western investment.
Meanwhile, Meta's AI systems account for roughly 16% of automated AI internet traffic globally — second only to OpenAI — reflecting just how much of the internet's AI activity now flows through Meta's infrastructure.
Where Meta is heading
Meta is pursuing two tracks simultaneously: keeping Llama alive as the world's leading open-weights model family, while building Muse Spark into a closed frontier product that can compete with the best models from OpenAI, Google, and Anthropic. Whether it can sustain both — and whether the open-weights community will follow it as it increasingly plays both sides — is the central question hanging over Meta's AI future.
