Meta Unveils Muse Spark: The First Model from Its Superintelligence Lab

Model Overview

Muse Spark is built on a completely rebuilt AI stack—new infrastructure, model architecture, and data pipelines—designed as a foundation for personal superintelligence that can perceive and reason about a user’s environment.

Performance

In multimodal perception, reasoning, medical‑related tasks, and various agent benchmarks, Muse Spark shows competitive capabilities. The “Contemplating” mode schedules multiple agents for parallel inference, allowing performance comparable to high‑intensity reasoning modes of Gemini Deep Think and GPT Pro. In the Humanity’s Last Exam test the model achieved 58% accuracy, and in the FrontierScience Research test it reached 38%.

Application Scenarios

Personal superintelligence enables AI to understand a user’s surroundings and assist in highly personalized tasks such as health management, interactive game generation, and device troubleshooting. Over 1,000 doctors contributed training data to improve medical‑reasoning accuracy, allowing the model to generate explanations of nutrition, muscle groups, and other health‑related information.

Scaling Axes

Pretraining

During a nine‑month overhaul the team redesigned model architecture, optimization methods, and data construction, achieving higher capability per unit of compute. Scaling‑law fits on a series of smaller models show that Muse Spark reaches the same performance as Llama 4 Maverick with more than an order of magnitude fewer FLOPs, indicating substantial efficiency gains.

Reinforcement Learning

After pretraining, reinforcement learning (RL) adds additional compute. Metrics such as pass@1 and pass@16 on training data grow logarithmically with RL steps, while test‑set accuracy rises steadily, demonstrating stable and predictable improvements without sacrificing reasoning diversity.

Test‑Time Reasoning

RL enables the model to “think” before answering. Token efficiency is optimized through a thinking‑time penalty and multi‑agent collaboration. AIME evaluations show that longer thinking initially improves results, then the penalty compresses inference so the model solves the same problems with fewer tokens. Adding parallel agents further enhances complex problem solving while keeping latency comparable to single‑agent extensions.

Demonstrations

Worked examples include:

Personalized nutrition recommendation with visual markers indicating recommended and non‑recommended foods, displaying health scores and macro‑nutrient information.

Identification of stretched muscle groups in yoga poses with difficulty ratings and corrective guidance.

Conversion of a static image into an interactive web‑based Sudoku game.

Component‑wise tutorial for operating a coffee‑machine, with hover‑activated highlights of key parts.

Availability

Muse Spark is currently available through Meta AI applications and a private API preview; pricing details have not been disclosed.

Reference URLs: https://ai.meta.com/blog/introducing-muse-spark-msl/ https://venturebeat.com/technology/goodbye-llama-meta-launches-new-proprietary-ai-model-muse-spark-first-since