Meta‘s new $299 glasses launched Tuesday with three frame styles, a celebrity collaboration, and an $80 price cut from its Ray-Ban predecessor. Running inside all of them is Muse Spark — the first proprietary AI model Meta has ever built, and the more consequential story of the two.
For a company that spent years building its AI reputation on open-source Llama releases, powering hundreds of third-party products along the way, shipping a closed-weight model is a significant strategic turn.
Muse Spark is the product of a full rebuild — the result of Zuckerberg cutting much of the Llama team after it fell significantly behind OpenAI, Google DeepMind, Anthropic, and xAI, and launching one of the most expensive AI talent acquisitions in the industry’s history. The company says it hopes to open-source future Muse models, a hedge that signals the decision wasn’t made lightly — and may not be permanent.
Sponsored · Altamira Gold Corp.
The architecture is built around a constraint every wearable AI faces — inference on a device you’re wearing has to be fast and efficient enough not to drain a battery before noon. Meta’s solution is a three-mode system — Instant, Thinking, and Contemplating — that adjusts response depth based on query complexity.
During training, a technique called thought compression rewards accuracy while penalizing the model for burning unnecessary compute, producing faster answers without sacrificing reasoning quality. Meta claims the result matches its previous best model, Llama 4 Maverick, at one-tenth the compute cost.
Muse Spark scores 52 at launch on the Artificial Analysis Intelligence Index — fourth globally, behind Google, OpenAI, and Anthropic’s leading models, but a significant leap from the 18 Llama 4 Maverick posted on the same benchmark.
Proprietary or not, what Muse Spark does in the meantime is lock a purpose-built AI into the world’s most widely distributed smart glasses. IDC puts Meta’s share of the smart glasses market at 69.2%, with Google’s Gemini-powered competitor still pre-launch. This means that every query users ask their glasses, every object identified, every translation run feeds a training loop no outside developer or competitor can access — a closed platform, deployed at scale, before the real competition arrives.
Privacy advocates aren’t taking Meta’s word for it — it wasn’t so long ago that Wired found that the company had quietly embedded a near-functional facial recognition system into the Meta AI companion app, installed on more than 50 million phones, without any public disclosure. The feature, internally called NameTag, was built using technology licensed from a firm that contracts with the US Marshals Service and Special Operations Command.
Meta stripped the code within a day of the story’s publication, then declined to answer whether it had already collected biometric data during internal testing. The company has not said the feature is permanently off the table.
