Mark Zuckerberg’s big bet on artificial intelligence is about to be put to the test. The CEO of Meta sent shockwaves through the AI landscape in 2025 when he spent millions to poach AI researchers from rival companies, assembling the star-studded team behind Meta Superintelligence Labs. After further restructuring and additional layoffs to balance out its spending spree, Meta’s AI transformation is settling into place, and the company finally has a new AI model to show for its efforts: Muse Spark.
What to Know About Meta’s Muse Spark
Muse Spark is the first AI model released by Meta Superintelligence Labs. It’s a compact model that lays the foundation for scaling up to larger models, but it still exhibits complex reasoning, multimodality and agentic capabilities. Muse Spark currently powers Meta’s AI assistant in the Meta AI app and meta.ai website, with plans to integrate it into the company’s smart glasses.
Muse Spark marks the beginning of the Muse model family and Zuckerberg’s official pursuit of “personal superintelligence,” according to a Meta blog post. While the model’s multimodality, agentic capabilities and complex reasoning offer a promising start, its consumer-first focus could reignite digital privacy concerns and put it at odds with the industry’s recent shift toward the enterprise space, leaving Meta with a ton of ground to make up in the AI race.
What Is Muse Spark?
Muse Spark is Meta’s newest model that powers the Meta AI assistant in the Meta AI app and meta.ai site. It’s smaller by design, with plans to scale up to “increasingly advanced models,” including some open-source options, according to a post by Zuckerberg. The company pre-trained Muse Spark by overhauling its tech stack and applying a scaling law to other small models for reference. This way, Muse Spark could develop advanced abilities in reasoning, understanding and coding, while requiring far less compute than its predecessor, Llama 4 Maverick.
Muse Spark also underwent reinforcement learning, or learning through trial and error in a controlled environment. This taught the model to reflect before answering questions, essentially thinking longer to master more complex reasoning. Muse Spark then used a process Meta refers to as “thought compression,” or minimizing the amount of tokens required to reason through a problem. As a result, plenty of compute is packed into this compact model, making it useful for a range of everyday applications.
What Can Muse Spark Do?
Here’s a closer look at what Muse Spark is capable of, from managing teams of agents to analyzing its surroundings.
Agent Orchestration
Muse Spark stands out for its ability to orchestrate multiple AI agents — systems that can perform complex, multi-step tasks without human assistance. When in “Contemplating” mode, Muse Spark can activate a team of these agents to perform tasks in parallel, solving more challenging problems without affecting the model’s performance or response time.
For instance, say a user submits a query about spending $300 or less on a flight to Denver, Colorado, and staying in an Airbnb near a ski resort and vegetarian dining options. One agent can look up flights under $300, another can map out various Airbnb locations in Denver, another can look up ski resorts in or near the city and another can look up vegetarian dining options. Operating simultaneously, these agents can provide more detailed insights, improving the accuracy and relevance of the model’s response.
According to Meta’s model evaluations, Muse Spark’s contemplating mode enables it to excel on the benchmark known as Humanity’s Last Exam, surpassing the performance of other industry heavyweights in Gemini 3.1 DeepThink and GPT 5.4 Pro.
Visual Understanding
Multimodal by nature, Muse Spark is designed to process visual information and understand the world around it. Users can simply take a photo of an object and ask Meta’s AI assistant questions about it, compare it to similar items and receive general information, getting answers in real time for a more interactive experience.
The model’s visual abilities extend into the digital world as well, making it adept at visual STEM problems and visual coding. Users can then enter prompts to Muse Spark to build websites, games, dashboards, simulations and other projects. Although the model can process visual, verbal and written queries, it can only produce text responses for now.
Health Reasoning
Muse Spark’s advanced reasoning makes it useful for addressing health-related questions specifically. According to Meta, the company partnered with more than 1,000 physicians to develop training data that supports “more factual and comprehensive responses.” This training process allows Muse Spark to provide guidance on exercise tips, different muscle groups and nutritional content, among other common health topics. The model can even enrich its answers with charts, images and other graphics to help users better understand its explanations.
Ethical Considerations
To account for Muse Spark’s STEM knowledge, Meta evaluated the model’s safeguards based on the “Advanced AI Scaling Framework,” implementing AI safety training and additional guardrails post-training. The company says the model consistently refuses unethical requests so far, particularly those involving biological and chemical weapons. That said, more research will be shared in Meta’s upcoming safety report.
Why Muse Spark Is a Risky Bet for Meta
In mid-2025, Zuckerberg presented his vision for personal superintelligence, arguing that artificial intelligence would be most impactful when supporting the needs and preferences of everyday consumers. Muse Spark reaffirms this consumer-first approach through its “Shopping mode,” which assists with styling for clothing and interior decorating, and its ability to surface public content posted by locals when a user searches a location.
As helpful as these features might be, individual consumers are unlikely to generate the same kind of revenue for Meta that enterprise customers could. Take Anthropic, for example. Its Claude chatbot is the go-to option for businesses about 73 percent of the time, which is a key reason why the company’s run-rate revenue has exceeded $30 billion in 2026. OpenAI has taken notice of its rival’s success, pivoting to an enterprise strategy and scrapping side projects that exhibited a consumer bent, most notably its Sora platform and “adult mode” for ChatGPT.
Many consumers may not want a personalized AI assistant, either. For instance, allowing Muse Spark users to submit health-related queries could lead to deeply personal conversations that compromise users’ digital privacy, especially if data is leaked or Meta shares its data with third parties. There’s also debate around whether these kinds of interactions could harm users’ well-being, posing challenges to consumer-first AI products.
Tech leaders’ embrace of the enterprise space and the privacy issues surrounding personalized AI beg the question of whether personal superintelligence can even be profitable. But it’s still early, with plenty of time for Meta to prove doubters wrong and become a meaningful player in the AI race once again.
What Does This Mean for the AI Race?
Artificial intelligence may be stealing headlines for its potential to upend the enterprise landscape, but it’s slowly making its way into people’s personal lives as well. The latest smartphones have incorporated more AI features as the hype around AI-first devices continues to build, and the next device to break onto the scene could be smart glasses. In fact, Meta plans to integrate Muse Spark into its own smart glasses, and it’s set to face direct competition from Apple, which also sees consumer-focused AI devices as a promising niche.
While none of these developments are necessarily game-changing, they have created enough excitement around the state of Meta’s AI products to send its Meta AI app climbing the App Store rankings. At the same time, Meta is expected to overtake Google as the company with the most net ad revenue in 2026, putting it in a favorable position to continue funding its AI initiatives for the long term.
All this to say that Meta isn’t going anywhere. On the contrary, it’s applying pressure on rivals like OpenAI and Anthropic who are banking on historic IPOs to establish additional revenue streams for their AI investments. And if Muse Spark-powered smart glasses become popular, Meta will have a leg up in the physical AI space as companies rush to build new devices that enable AI to better engage with its environment.
Muse Spark may not be on the same level as top models like Gemini 3 or Claude Code, but it’s a step toward designing larger models that could one day justify Meta’s unprecedented reorganization around AI and propel the company to the forefront of the AI race.
Frequently Asked Questions
What is Muse Spark?
Muse Spark is the first AI model released by Meta’s Superintelligence Labs, possessing complex reasoning, multimodal understanding and agentic capabilities despite its smaller size. The model currently powers Meta’s AI assistant in the Meta AI app and meta.ai site, and it could soon be incorporated into Meta’s smart glasses.
Is Muse Spark free?
Yes, users can access Muse Spark for free on the meta.ai website and the Meta AI mobile app.
How does Muse Spark stack up against other AI models?
Given its compact size, Muse Spark isn’t quite on the same level as larger models like Google’s Gemini 3, Anthropic’s Claude or OpenAI’s GPT-5. But it still demonstrates impressive reasoning, holding its own against Gemini 3.1 and GPT 5.4 Pro on Humanity’s Last Exam. Muse Spark also outperforms all other models in successfully refusing queries involving biochemical weapons, according to a Meta blog post.
Why is Meta’s AI strategy risky?
Meta’s AI strategy revolves around CEO Mark Zuckerberg’s pursuit of “personal superintelligence,” which focuses on building AI that caters to individual consumers’ preferences. This approach conflicts with the industry’s more recent shift toward enterprise AI, which has been spearheaded by Anthropic and OpenAI. As a result, prioritizing personalized AI may put Meta at a disadvantage compared to competitors who rely on enterprise customers as a major revenue stream.
