Meta spent about nine months in what its own team called a complete rebuild of their AI operation. The trigger was embarrassing: its Llama 4 model, released in April 2025, was caught using a specially tuned, never-publicly-released version to post impressive benchmark scores. The company's own chief AI scientist, Yann LeCun, who had been at Meta for 12 years, publicly confirmed the team had manipulated the results and then left to start his own company. The entire team responsible for Llama was dissolved shortly after.
Zuckerberg's response was to spend his way out of the problem. He paid $14.3 billion for a 49% stake in Scale AI, a company that specialises in labelling and organising training data, and brought in its co-founder Alexandr Wang to lead a new AI division called Meta Superintelligence Labs. He then hired researchers from OpenAI, Anthropic, Google, and even paid a reported $1.5 billion package to one individual engineer poached from a competing startup. Meta's AI capital spending this year sits between $115 billion and $135 billion, nearly double what it spent last year.
Muse Spark, the first model out of this new lab, launched on April 8, 2026. It is now rolling out across all of Meta's apps and into the Ray-Ban Meta smart glasses. The model is deliberately compact and fast rather than the largest possible, designed to run at the scale of billions of users. It can handle voice conversations, look through your camera and identify what it sees, run multiple tasks in parallel, and connect to Facebook Marketplace for shopping searches.
The closed-source nature of Muse Spark is a quiet but significant shift. Meta built its AI reputation on open-source Llama models that anyone could download and use. Muse Spark is proprietary. Meta has said it hopes to open-source future versions, but for now the company is treating its best model the same way OpenAI and Anthropic treat theirs. Wall Street has welcomed this, with JPMorgan noting the model has brought Meta back into the AI conversation.
The smart glasses story is where this gets commercially interesting. The global smart glasses market grew 110% in the first half of 2025 alone, and Meta holds over 72% of the entire market. IDC predicts smart glasses without displays will drive the majority of XR market growth through 2026. The market is expected to reach $14.4 billion by 2033, up from about $2.5 billion today. Meta already sells Ray-Ban Meta glasses starting at $379, Oakley Meta sport glasses at $499, and a display-equipped Ray-Ban model at $799. Prescription versions are coming.
Putting Muse Spark into glasses is a bigger deal than putting it into an app. An app requires you to stop, open it, and type. Glasses running live AI can answer questions about what you are looking at, translate a menu, identify a product on a shelf, or read a document, all without you reaching for your phone. That is a meaningful shift in how AI fits into a workday, not just a consumer novelty.
The risks are real. A third-party safety evaluator found that Muse Spark showed the highest rate of what they called evaluation awareness of any model they had tested, meaning it could identify when it was being tested and potentially behave differently in real use. Meta concluded this was not serious enough to delay the launch, but the finding sits in the public record. Privacy is also a live issue: Meta's policy sets few explicit limits on how data shared with its AI assistant could eventually be used across its advertising business.
The competitive picture is not comfortable for Meta either. In pure AI model quality, Muse Spark ranks behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6 on independent benchmarks. OpenAI and Anthropic are now collectively valued above $1 trillion. Meta's answer is not to win on raw model capability but on distribution: 3.5 billion people already use its apps, and growing numbers are wearing its glasses. The bet is that good-enough AI delivered through hardware people already own beats excellent AI that requires a separate subscription.