Meta Launches Muse Spark to Power Personal Superintelligence Across Social Apps

Meta Unveils Muse Spark

Meta today announced the release of Muse Spark, the debut large language model from its newly formed Meta Superintelligence Labs. This launch signals a major shift toward proprietary software, moving away from the open source heritage of the Llama series. Muse Spark is now powering Meta AI on the web and will roll out to Instagram, Facebook, and WhatsApp in the coming weeks. The model features native multimodality, allowing it to see and reason through dynamic environments using a visual chain of thought. Users can snap photos of real world objects to receive instant analysis, such as identifying protein content in snacks or comparing product labels. This capability is expected to be a cornerstone for Meta’s upcoming AI glasses, where the assistant can see the world alongside the user.

Advanced Reasoning and Multi-Agent Tasks

A new Contemplating mode enables the AI to launch multiple sub-agents in parallel to solve complex problems. For example, one agent can plan a flight itinerary while others research local activities and weather simultaneously. This multi-agent approach aims to deliver faster and more comprehensive answers for high level reasoning tasks that require cross-referencing multiple data points at once. Developed over nine months under the leadership of Chief AI Officer Alexandr Wang, the model utilizes a technique called thought compression to maintain high accuracy with less compute. While it trails slightly behind competitors in pure software engineering, it ranks as one of the top vision models globally. It also includes specialized training for health queries, developed in collaboration with over 1,000 physicians to help users interpret medical charts or common health concerns.

Ecosystem Integration

Beyond chat, Muse Spark introduces a shopping mode that leverages social trends and creator content to offer personalized style recommendations. It also supports visual coding, enabling users to generate custom mini-games and dashboards directly from text prompts within the Meta AI app. Meta indicated that while this first version is proprietary, the company may open source future iterations as the Superintelligence Labs expand development.