If you’ve used Meta AI on your phone or through the meta.ai website in the past few days, the model behind it has quietly changed. Meta began rolling out Muse Spark this week across the Meta AI app and meta.ai in the United States, marking the first product shipped by the company’s newly formed Superintelligence Labs. The release follows a sweeping reorganization of Meta’s AI operations, anchored by a $14.3 billion investment in Scale AI and the recruitment of Scale AI CEO Alexandr Wang to lead the superintelligence effort.
Together, these moves amount to Meta’s clearest bet yet on a consumer-first AI strategy, one built on outside talent, massive capital, and a model the company says was “purpose-built to prioritize people.”
What Muse Spark actually does
Meta describes Muse Spark as a multimodal model with three distinct operating modes. A fast mode handles quick, conversational queries. A reasoning mode tackles more complex, multi-step problems. And a shopping mode tailors product recommendations to individual user interests and browsing behavior across Meta’s apps, a feature that could reshape how hundreds of millions of people discover and buy things online.
For now, the new features are live only in the United States. Meta says expansion to WhatsApp, Instagram, Facebook, Messenger, and Meta AI glasses will follow “in the next few weeks,” though no specific dates or international markets have been confirmed. A private-preview API is also open to select developer partners.
The deal and the team behind it
The organizational backstory matters because it explains why this launch looks different from Meta’s previous AI releases. In a deal reported by Bloomberg, Meta acquired a 49% stake in Scale AI for $14.3 billion, then brought Wang on to run its Superintelligence Labs. Wang continues to serve as Scale AI’s CEO, though the precise governance arrangement between the two organizations has not been detailed in public filings or press materials.
That dual role raises questions investors and developers are already asking: How do Meta and Scale AI share data, infrastructure, and intellectual property? What authority does Wang hold over product decisions at Meta? And how does Scale AI’s core business in data labeling feed into Meta’s model training pipeline? None of these questions have been publicly answered.
It is also worth noting that Muse Spark is the first model from this specific unit, not Meta’s first AI model. The company has shipped multiple generations of its open-source Llama models under different organizational structures. Whether Muse Spark will follow the same open-source approach or remain proprietary is another unanswered question.
The health claims deserve scrutiny
In press materials, Meta has pointed to the HealthBench benchmark to support claims about Muse Spark’s performance on health-related queries. HealthBench is a legitimate, peer-reviewed evaluation framework for large language models, with published methodology, datasets, and rubrics that allow independent testing. The benchmark and its more demanding variant, HealthBench Hard, are not in dispute as rigorous tools.
What is missing: Meta has not published exact scores, disclosed which version of the benchmark it used, or provided a reproducible scorecard. The arXiv paper defines the standard. It does not validate Meta’s specific claims. Until Meta or an independent lab publishes verifiable results, the health-AI angle rests entirely on the company’s word.
That gap matters more than it might for a typical product launch. Health advice and purchasing recommendations are areas where accuracy carries real consequences. A model that confidently delivers wrong medical information or steers users toward products based on opaque behavioral profiling is not just a bad product; it is a liability.
How this stacks up
Meta has not published head-to-head comparisons between Muse Spark and competing models from OpenAI, Google, or Anthropic. Without independent benchmarks, it is difficult to say where Muse Spark sits in the current landscape. What is clear is that Meta is taking a different strategic path: rather than selling API access or charging for a premium chatbot, it is embedding its most advanced model directly into apps that billions of people already use daily.
That distribution advantage is significant. OpenAI’s ChatGPT and Google’s Gemini require users to seek them out. Muse Spark, by contrast, will surface inside WhatsApp conversations, Instagram feeds, and Facebook Messenger threads, meeting users where they already spend their time. Whether the model’s quality justifies that privileged placement is the question Meta will have to answer over the coming months.
What to watch
For anyone already using Meta AI in the U.S., the switch to Muse Spark is already live. If the shopping mode works as described, expect product recommendations that increasingly reflect your behavior across Meta’s platforms. Users who rely on Meta AI for health-related questions should treat its responses with the same caution they would apply to any AI tool, especially given the absence of published benchmark results.
The bigger story will unfold over the next several weeks as Muse Spark reaches WhatsApp, Instagram, and Meta’s other properties. How quickly international markets get access, whether independent researchers can reproduce Meta’s performance claims, and how Wang’s dual role at Meta and Scale AI evolves will determine whether this launch is remembered as a genuine strategic shift or an expensive rebranding of capabilities Meta already had.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
