Meta Launches Muse Spark: Connects the Social Media Ecosystem—An AI Health and Shopping Assistant Just for You

Meta’s superintelligence lab officially released its first in-house AI model, Muse Spark, on April 8. Positioned as the first step toward “personal superintelligence,” the news sparked a more than 6.5% jump in Meta’s stock price in a single day, hitting a three-week high, and also showed that Meta hasn’t lost momentum in the AI race.

What is Muse Spark? Meta’s first in-house AI reasoning model goes live

Muse Spark is the first model in the “Muse series” developed by Meta’s Superintelligence Lab (Meta Superintelligence Labs, MSL). It is a native multimodal reasoning model that supports tool use (tool-use), a visual chain of thought, and multi-agent orchestration.

The lab is led by Alexandr Wang, the founder of Scale AI. Meta previously invested about $15 billion in Scale AI and recruited Wang, which was seen as a key step toward rebooting its AI strategy. The release of Muse Spark is the first major result following CEO Mark Zuckerberg’s push to transform into AI in recent years.

Notably, unlike Meta’s past focus on open-source Llama models, Muse Spark uses a closed-source model, signaling Meta’s shift from an open ecosystem strategy toward a more commercialization-driven AI development path. Market analysis suggests that in the future, it is not out of the question to roll out paid APIs or subscription-based services.

As of now, Muse Spark is available to use on Meta AI, and it also offers a private API preview to select partners.

Muse Spark performance: benchmark against Gemini Deep Think and GPT Pro

In terms of capability evaluation, Muse Spark demonstrates strong performance in multimodal perception, reasoning, health information processing, and agent tasks. Through the “Contemplating” mode introduced by Meta, multiple agents perform parallel reasoning and collaborate, enabling Muse Spark to stand alongside extreme reasoning modes such as Gemini Deep Think and GPT Pro on difficult tasks.

Muse Spark achieved 58% in the Humanity’s Last Exam test and 38% in the FrontierScience Research test, demonstrating实力 that can compete with top-tier models.

Use cases: from health tracking to a shopping assistant, fully integrating Meta’s ecosystem

Muse Spark is positioned as personal superintelligence that “understands the world,” emphasizing multimodal visual integration to analyze users’ real-time environments across domains. It includes the following applications in particular.

Multimodal interaction

Muse Spark can generate interactive web mini-games based on images, or recognize objects and provide detailed explanations—for example, parsing complex machinery operation tutorials.

Personal health

The model can combine visual recognition with tool search to provide more precise health information responses—for example, analyzing the nutritional components of food or indicating how different muscle groups are active during exercise. It also provides personalized suggestions based on individual dietary restrictions (such as vegetarians or people with high cholesterol).

Meta has previewed that it will gradually integrate Muse Spark into platforms including Instagram, Facebook, Messenger, WhatsApp, and smart glasses. Among them, the AI shopping assistant is seen as a key commercial application. It will help users search for products, provide recommendations, and support decision-making—showing that the outlines of a business model combining ads and e-commerce have been taking shape step by step.

Three core technical pillars: pretraining, reinforcement learning, and test-time reasoning

Meta revealed the technical core of Muse Spark in its official blog, building around three expansion pillars.

Pretraining

Over the past nine months, Meta rebuilt its pretraining technology stack, upgrading the model architecture, optimization methods, and data preparation comprehensively, significantly improving the model capabilities extractable per unit of compute.

Reinforcement Learning

RL, as an extension of pretraining, can expand model capability at larger scale. Meta said that as RL compute (measured by number of steps) scales up, the probability of success on training data increases in a log-linear fashion. On evaluation efficiency for unseen datasets, it likewise shows stable improvement.

Test-Time Reasoning

RL training makes the model think before answering. Meta, through a “thought compression” mechanism and multi-agent collaboration, effectively reduces latency and strengthens reasoning ability.

The curtain rises for the era of personal superintelligence, and investors are optimistic that AI will boost user stickiness

After the news broke, Meta’s stock rose 6.5% to close at $612.42 yesterday, setting a three-week high. This indicates that investors broadly believe in applying AI capabilities to everyday personal tasks and combining them with social platform potential.

Meta said that Muse Spark represents a key step toward the company’s “predictable and high-efficiency AI scaling trajectory.” In the future, it will continue to release models with even stronger capabilities, steadily moving toward its goal of personal superintelligence.

The article Meta launches Muse Spark: connecting the social media ecosystem and becoming a dedicated AI health and shopping assistant first appeared on Liannews ABMedia.