Meta’s Muse Spark model
Meta introduced Muse Spark, described as its first post‑AI‑reboot model with competitive benchmarks, marking another public iteration in Meta’s model roadmap. Coverage frames Muse Spark as part of Meta’s broader push to combine open and closed approaches in its AI stack. (x.com/WIRED)
Meta introduced Muse Spark on April 8 as the first model from Meta Superintelligence Labs, and it now powers the Meta AI app and meta.ai. (about.fb.com) Muse Spark is a large language model, the kind of system that predicts the next word in a sequence, but Meta says this one was built to handle both quick replies and harder reasoning tasks. Meta said the model is “small and fast by design,” with support for science, math, health, and image-based prompts. (about.fb.com) Meta said users can switch between modes in the Meta AI app, and the assistant can run multiple subagents in parallel on one request. In Meta’s example, separate agents can draft a Florida itinerary, compare Orlando with the Florida Keys, and find kid-friendly activities at the same time. (about.fb.com) The company tied Muse Spark to products it already controls. Meta said the model will roll out in the coming weeks to WhatsApp, Instagram, Facebook, Messenger, and its artificial-intelligence glasses, and it plans a private-preview application programming interface for selected partners. (about.fb.com) That marks a shift from Meta’s recent public framing around Llama, its open-weight model family. In April 2025, Meta launched the standalone Meta AI app “built with Llama 4”; with Muse Spark, Meta is now putting a new in-house model directly inside its consumer products while keeping outside access limited to selected partners. (about.fb.com, about.fb.com) Meta said Superintelligence Labs rebuilt its artificial-intelligence stack “from the ground up” over the last nine months. Alexandr Wang, the former Scale AI chief executive, joined Meta in June 2025 as its first chief artificial intelligence officer and now leads that unit. (about.fb.com, meta.com) Outside coverage has described Muse Spark as a proprietary model rather than another open release. VentureBeat reported that Meta said Muse Spark reaches its reasoning performance with more than an order of magnitude less compute than Llama 4 Maverick, Meta’s previous mid-size flagship. (venturebeat.com, about.fb.com) Meta is also using Muse Spark to pull its social platforms closer to its assistant. The company said future features will cite recommendations and content shared across Instagram, Facebook, and Threads, which would make Meta AI more dependent on information flowing through Meta’s own apps. (about.fb.com) The immediate test is whether Muse Spark can make Meta AI feel faster and more useful across the company’s biggest products, not just on benchmark charts. Meta said larger Muse models are already in development, which makes Spark the first public checkpoint in that next model line. (about.fb.com)
