Meta rolls out Muse Spark, 'Contemplating' mode
Meta has unveiled a model called Muse Spark and started rolling out a free "Contemplating mode" that toggles advanced multi‑agent reasoning in the Meta AI app, while analysts flag a much larger capital spending push tied to personal superintelligence efforts. Coverage frames the technical launches alongside discussion of substantial capex commitments behind them. ( )
Meta has started rolling out “Contemplating” mode in the Meta AI app, adding a deeper reasoning setting to the new Muse Spark model it introduced on April 8. (about.fb.com) (testingcatalog.com) Muse Spark is the first model in Meta’s new Muse series from Meta Superintelligence Labs, and Meta said it already powers the Meta AI app and meta.ai. The company said it will reach WhatsApp, Instagram, Facebook, Messenger and its artificial intelligence glasses “in the coming weeks,” and a private application programming interface preview is planned for selected partners. (about.fb.com) Meta describes the new setting in plain terms as parallel problem-solving: the assistant can launch multiple subagents at once, with one drafting, another comparing options and a third gathering supporting details before combining the answer. Meta used a family trip to Florida as its example of how the mode splits one prompt into several tasks. (about.fb.com) The product push sits inside a larger reset of Meta’s artificial intelligence strategy. Meta said Meta Superintelligence Labs rebuilt its artificial intelligence stack over the last nine months, and Chief Executive Officer Mark Zuckerberg tied 2026 to “personal superintelligence,” his term for a more individualized assistant. (about.fb.com) (investor.atmeta.com) (about.fb.com) That strategy also carries a much larger infrastructure bill. Meta told investors on January 28 that 2026 capital expenditures would run between $115 billion and $135 billion, up from $72.22 billion in 2025, with spending tied to Meta Superintelligence Labs efforts and the core business. (investor.atmeta.com) (datacenterdynamics.com) Meta has been laying out where that money goes. In March, it said four new generations of its Meta Training and Inference Accelerator chips were due within two years, and in February it announced a long-term Advanced Micro Devices deal for up to 6 gigawatts of Instinct graphics processing units for its artificial intelligence infrastructure. (about.fb.com 1) (about.fb.com 2) Muse Spark also marks a product change from the Meta AI app Meta launched on April 29, 2025, when the company said the standalone assistant was built with Llama 4. The new April 2026 announcement no longer centers Llama for the app and instead positions Muse Spark as the model behind Meta AI’s current upgrade. (about.fb.com 1) (about.fb.com 2) Meta says Muse Spark is small and fast by design, but built for harder reasoning and multimodal work, meaning it can take text, images and voice input instead of only typed prompts. The company said the model will eventually support features that cite recommendations and content shared across Instagram, Facebook and Threads. (about.fb.com) For users, the immediate change is simple: Meta AI is becoming a two-speed assistant, with a standard mode for quick replies and “Contemplating” mode for slower, heavier questions. For investors, the same launch lands as an early product test of whether Meta’s 2026 spending surge can turn a social app audience into a market for paid-up artificial intelligence infrastructure. (about.fb.com) (investor.atmeta.com)