Meta has unveiled Muse Spark, the first model in a new large language model series developed by Meta Superintelligence Labs (MSL). The model now powers the Meta AI assistant across the Meta AI app and meta.ai, with a rollout to WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses expected in the coming weeks.

A New Model Architecture

Built over nine months from a rebuilt AI stack, Muse Spark is designed to be small and fast while capable of reasoning through complex questions in science, mathematics, and health. It is the first in what Meta describes as a deliberate, scientific model scaling series — each generation intended to validate findings before larger models are developed. Meta confirmed that a next-generation model is already in development.

The model supports both an Instant mode for quick responses and a Thinking mode for complex reasoning tasks. Meta AI can now also launch multiple subagents in parallel, enabling it to handle multi-part queries simultaneously.

Multimodal and Health Capabilities

Muse Spark includes strong multimodal perception, allowing Meta AI to interpret images and visual inputs rather than relying solely on text. The company worked with a team of physicians to develop the model’s ability to handle health-related questions, including those involving charts and images. Visual coding capabilities are also included, enabling users to generate custom websites and interactive applications directly from prompts.

Shopping and Social Context Integration

A new shopping mode allows Meta AI to surface product recommendations drawn from content and creator activity across Instagram, Facebook, and Threads. The feature is designed to connect purchasing decisions with existing social signals rather than external search. Similarly, Meta AI can now surface posts from local communities when users ask about specific locations or trending topics.

API Access and Open Source Plans

Meta will offer Muse Spark in private preview via API to select partners. The company also indicated it hopes to open-source future versions of the model. The initial feature rollout is starting in the United States, with broader country availability and platform expansion to follow.