Meta shifts model strategy

A year ago, Meta was telling developers to build on Llama 4. This week, it introduced Muse Spark, said it already powers the Meta Artificial Intelligence app and website, and said WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses will get it in the coming weeks. (about.fb.com 1) (about.fb.com 2) That is a sharp change in how Meta is packaging its artificial intelligence. Llama 4 was released with downloadable model weights through llama.com and Hugging Face, while Muse Spark is being offered only inside Meta products and in private preview to selected partners through an application programming interface. (about.fb.com 1) (about.fb.com 2) (about.fb.com 3) Meta is not hiding the handoff. Its April 8 announcement called Muse Spark “the first in a new series” from Meta Superintelligence Labs, and TechCrunch reported the launch as the first step in a broader overhaul after frustration inside the company with how Llama stacked up against ChatGPT and Claude. (about.fb.com) (techcrunch.com) The practical difference is simple: an open-weight model is like selling the engine, while a closed model is like selling rides in the car. With Llama, outside developers could download the weights and run the model on their own systems; with Muse Spark, Meta keeps the model behind its own walls and decides who gets access. (about.fb.com 1) (about.fb.com 2) That gives Meta tighter control over speed, safety settings, and product design across its apps, which now reach billions of people. It also means the company can improve one central assistant and push those changes into Instagram direct messages, Facebook search boxes, Messenger chats, and Ray-Ban smart glasses without waiting for an outside ecosystem to catch up. (about.fb.com) (about.fb.com) The timing is awkward because Meta is also losing room to argue that it is just a neutral platform. On April 10, the Massachusetts Supreme Judicial Court said the state can keep pursuing claims that Instagram was designed to addict young users, rejecting Meta’s attempt to block the case under the federal law often used to shield internet companies from liability over user content. (bloomberglaw.com) (aol.com) That ruling focuses on product design, not just what people post. The Massachusetts attorney general’s case argues Meta built features that keep teenagers engaged for longer stretches, and the court let that theory move forward. (aol.com) (nypost.com) California added another hit two weeks earlier. On March 25, a Los Angeles County jury found Meta and Google liable in a teen mental-distress case, and California Courts said the plaintiff and her mother were awarded $3 million, with Meta responsible for about 70 percent. (newsroom.courts.ca.gov) After that verdict, Meta said on April 9 that it was pulling ads from Facebook and Instagram that had been used by law firms to recruit more plaintiffs for social-media-addiction cases. Reuters reported the company removed those ads while the broader litigation keeps moving. (reuters.com) (axios.com) Put together, the picture is not just “new model replaces old model.” Meta is concentrating more of its artificial intelligence inside products it controls at the same moment courts in Massachusetts and California are looking more closely at how those products are designed and what those design choices do to young users. (about.fb.com) (bloomberglaw.com) (newsroom.courts.ca.gov)