Muse Spark: Scaling towards personal superintelligence

Meta Superintelligence Labs has unveiled Muse Spark, the inaugural model in its Muse family and the first product from a complete overhaul of Meta's AI strategy. This natively multimodal reasoning model supports tool-use, visual chain-of-thought, and multi-agent orchestration, representing Meta's foundational step toward developing personal superintelligence. Available immediately through meta.ai and the Meta AI app (with private API previews for select users), Muse Spark demonstrates competitive performance across multimodal perception, reasoning, health, and agentic tasks.

A standout feature is 'Contemplating mode,' which orchestrates multiple reasoning agents in parallel to compete with extreme reasoning modes from frontier models like Gemini Deep Think and GPT Pro. This approach achieved 58% on the Humanity's Last Exam benchmark and 38% on FrontierScience Research. The model is built from the ground up for visual integration, enabling practical applications like creating interactive web games from prompts or generating annotated tutorials for home appliances.

For health applications—a major focus for personal superintelligence—Meta collaborated with over 1,000 physicians to curate training data. This enables Muse Spark to generate interactive displays explaining nutritional content, localize recommended foods for specific diets, and analyze exercise form by highlighting activated muscles. The model can produce personalized health scores and justifications in real-time visualizations.

Meta is supporting this launch with strategic investments across its entire AI stack, including the Hyperion data center infrastructure, to enable further scaling. The company outlines three key scaling axes: pretraining, reinforcement learning, and test-time reasoning, positioning Muse Spark as the first predictable step toward models that understand and interact with a user's personal world.