Meta pushes Muse Spark

Meta has begun rolling out Muse Spark across its own products, shifting its artificial intelligence push away from Llama and toward a model built first for Meta’s apps and devices. (about.fb.com) Meta announced Muse Spark on April 8, 2026, saying the model already powers the Meta AI app and website and will expand to WhatsApp, Instagram, Facebook, Messenger and its artificial intelligence glasses in the coming weeks. Meta also said it will offer the model in private preview through an application programming interface to select partners. (about.fb.com) The company described Muse Spark as the first model from Meta Superintelligence Labs, a unit it said rebuilt Meta’s artificial intelligence stack over the last nine months. Meta said the model is “small and fast by design,” but built for reasoning and for handling text, images and voice-driven tasks inside consumer products. (about.fb.com) That marks a break from Meta’s recent pattern with Llama, which centered on open-weight releases aimed at outside developers. Artificial Analysis said Muse Spark is Meta’s first frontier-class model not released as open weights and the company’s first new major release since Llama 4 in April 2025. (artificialanalysis.ai) Meta is tying the new model to a broader plan to make its assistant more personal by using context from its own network of apps. On Meta’s January 28, 2026 earnings call, Chief Executive Mark Zuckerberg said the company was building systems that understand a user’s history, interests, content and relationships, and merging large language models with the recommendation systems behind Facebook, Instagram, Threads and ads. (s21.q4cdn.com) The spending behind that plan is unusually large even by Meta standards. In its full-year 2025 results, Meta said capital expenditures reached $72.22 billion in 2025, and the company told investors it expects 2026 capital spending of $115 billion to $135 billion, with growth driven by infrastructure and Meta Superintelligence Labs. (nasdaq.com, quartr.com) Meta’s pitch is that distribution is the advantage: more than 3.5 billion people used at least one of its apps every day at the end of 2025, Zuckerberg said on that same call. Muse Spark is being positioned to ride that audience through built-in placement rather than a broad public model release on day one. (s21.q4cdn.com, about.fb.com) Early outside benchmarking suggests the model is competitive, though not clearly at the top of the field. Artificial Analysis said Muse Spark scored 52 on its Intelligence Index, placing it behind Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6, while outperforming several smaller or mid-tier rivals. (artificialanalysis.ai) Meta is also limiting how much of Muse Spark reaches outsiders for now. The company said the model is currently meant for Meta AI and other first-party services first, with only a private-preview application programming interface for selected partners rather than a broad developer release. (about.fb.com) The immediate test is not just whether Muse Spark can score well on benchmarks, but whether Meta can turn a rebuilt model stack and a bigger infrastructure bill into features people actually use inside Facebook, Instagram, WhatsApp and its glasses. Meta’s rollout schedule puts that test in front of users over the next few weeks. (about.fb.com, artificialanalysis.ai)