Meta launches Muse Spark AI model for multimodal reasoning and orchestration

Meta has introduced Muse Spark, its inaugural major AI model developed by the Meta Superintelligence Labs, which is spearheaded by Alexandr Wang. This innovative model excels in multimodal reasoning, enabling it to process various data types such as text and images. According to Meta, Muse Spark marks the beginning of a transformative journey in their AI initiatives. The company described it as a foundational product from a comprehensive overhaul of their AI capabilities. The model is engineered to tackle problems systematically, leveraging tools to visually present its reasoning while concurrently managing multiple AI systems. With its launch, Muse Spark is positioned as the first model under the Superintelligence Labs, a division established by CEO Mark Zuckerberg in July 2025. This groundbreaking model aims to provide “personal intelligence” for daily tasks, including visual comprehension, health management, shopping experiences, and social media interactions. Zuckerberg emphasized in a Facebook post that Meta is focused on creating products that not only respond to inquiries but also act autonomously on behalf of users. Currently available through the web and the Meta AI application, Muse Spark is designed for continuous improvement. As users engage with it, the model learns and enhances its capabilities. One notable feature is the “Contemplating mode,” which allows the AI to utilize multiple agents simultaneously to dissect complex issues more effectively. In terms of performance, Muse Spark has demonstrated impressive results, scoring 58.4% on Humanity’s Last Exam and 38.3% on Frontier Science Research. Additionally, it showcases health reasoning abilities developed with the input of over 1,000 healthcare professionals, providing insights on nutrition and exercise. Benchmark testing reveals that Muse Spark achieved 42.1% on HealthBench Hard, surpassing GPT 5.4 and outpacing other models like Gemini 3.1 Pro. On DeepSearchQA, it scored 74.8%, maintaining a competitive edge yet trailing slightly behind the top performers. It also reached 77.4% on SWE-Bench Verified, a figure close to leading scores in the low 80s, while achieving 52.4% on the more demanding SWE-Bench Pro, just shy of the best score of 57.7. Looking forward, Zuckerberg announced plans for more advanced models that will push the boundaries of intelligence and functionality, including new open-source offerings. The overarching goal is to develop tools that do more than just answer questions—they will actively assist users in their daily lives.