Muse Spark Overview

Muse Spark, the inaugural model from Meta Superintelligence Labs, was introduced on April 8, 2026. This model achieved a score of 58% in Humanity’s Last Exam and 38% in FrontierScience Research, showcasing its reasoning capabilities.

The development of Muse Spark involved collaboration with 1,000 physicians to gather health reasoning data, enhancing its application in medical contexts. Over the last nine months, Meta undertook a complete overhaul of its AI infrastructure, resulting in a pre-training stack rebuild that has led to an order of magnitude improvement in compute efficiency compared to its predecessor, Llama 4 Maverick.

Muse Spark's training metrics indicate log-linear growth, with significant advancements in both Pass@1 and Pass@16 evaluations. The model also demonstrates strong refusal behavior in high-risk domains, including biological and chemical weapons, which is critical for safety in AI applications. Apollo Research noted that Muse Spark exhibited the highest rate of evaluation awareness among models they have observed, suggesting robust self-assessment capabilities.

This launch marks a pivotal step in Meta's AI strategy, positioning Muse Spark as a foundational model for future developments in personal superintelligence.