Meta Reenters AI Race with Muse Spark

Meta Superintelligence Labs introduced Muse Spark, the first model in a new series positioned as Llama’s successor and designed to power Meta’s AI experiences across its product suite. Muse Spark already runs the Meta AI app and website in the U.S.; Meta says it will appear in WhatsApp, Instagram, Facebook, Messenger, Meta smart glasses, and additional countries in the coming weeks. The model will also be available to select partners via a private-preview API.

Muse Spark is described as a multimodal model supporting text and image inputs and capable of orchestrating multiple AI sub-agents to handle queries “better and faster.” It offers a user-facing speed/quality tradeoff with “Instant” and “Thinking” modes, mirroring recent product-level choices from competitors like Google Gemini. Meta highlights Muse Spark’s ability to answer complex questions in science, math, and health — a capability set that requires improved reasoning, chain-of-thought handling, and domain-aware retrieval or tool use in production.

Meta frames Muse Spark as “purpose-built for Meta’s products,” emphasizing integration across high-scale social and messaging surfaces and specialized hardware (camera glasses). The sub-agent architecture suggests internal modularization for parallel task execution or agent specialization. The private preview API indicates Meta intends limited external developer access initially rather than an immediate broad public model release.

Muse Spark represents Meta shifting from standalone model publications to product-integrated model deployments at scale. For ML engineers, the model’s support for multimodal inputs and sub-agent orchestration implies engineering investments in routing, latency controls, safety filters, and tool/knowledge integration. Product designers and infra teams should expect tradeoffs between latency and depth (Instant vs Thinking modes), and higher operational demand from embedding multimodal inference into chat, camera, and AR hardware.

verification of Muse Spark’s claimed capabilities on reasoning and domain-specific queries, the latency/throughput profiles of the sub-agent approach, safety and hallucination mitigation strategies especially for health queries, and the scope and terms of the private-preview API for partners.