Muse Spark: Meta’s Groundbreaking AI Model

Muse Spark is the latest significant AI model from Meta, created by the Meta Superintelligence Labs team headed by Alexandr Wang. This model demonstrates multimodal reasoning, enabling it to process various types of data, including both text and images.

Meta announced in a blog that Muse Spark represents the initial phase of their ambitious plans for AI advancements, marking a comprehensive update to their AI initiatives.

Muse Spark: Features and Performance

Muse Spark is the inaugural model launched under the Meta Superintelligence Labs, a division established by CEO Mark Zuckerberg in July 2025. This innovative model aims to provide personal intelligence tailored for daily tasks, including visual comprehension, health management, shopping, and social media interactions.

In a Facebook post, Zuckerberg shared that Meta is dedicated to creating solutions that not only respond to inquiries, but also function as intelligent agents performing tasks on behalf of users. The Muse Spark AI model is currently accessible through the web and the Meta AI application, and it is designed to evolve by learning from user interactions, thereby enhancing its capabilities over time.

Contemplating Mode

The AI model incorporates a ‘Contemplating mode’, enabling it to deploy multiple agents concurrently to analyse problems. These agents operate in parallel, enhancing the model’s reasoning efficiency, making it more adept at tackling intricate challenges. Muse Spark has demonstrated a 58.4% success rate in Humanity’s Last Exam and 38.3% in Frontier Science Research.

Health Reasoning Capabilities

Muse Spark also boasts health reasoning functionalities, developed in collaboration with over 1,000 healthcare professionals contributing to its training data. This feature can provide users with valuable health insights, such as nutritional information about foods and details about muscles activated during physical activities.

Benchmark Performance

When considering benchmark results, Muse Spark achieved a score of 42.1% on HealthBench Hard, slightly surpassing GPT 5.4, while significantly outperforming other models like Gemini 3.1 Pro. On DeepSearchQA, it recorded a score of 74.8%, positioning it ahead of several competitors but slightly behind the leading models. In SWE-Bench Verified, it attained 77.4%, nearing the top scores within the low 80s range. For the more demanding SWE-Bench Pro, Muse Spark scored 52.4, again trailing behind the top score of 57.7.

Looking to the future, Zuckerberg emphasised plans for increasingly sophisticated models designed to expand the boundaries of artificial intelligence and capabilities, including the introduction of new open source models. They are committed to advancing products that act as intelligent agents fulfilling various tasks for users.