Mark Zuckerberg announces Muse Spark, a new Meta AI model: How to try it, benchmark results

Mark Zuckerberg announced Wednesday that Meta Superintelligence Labs has reached its first major milestone: a new family of AI models called Muse, with the debut model, Spark, available now. In a Facebook post, Zuckerberg said that Muse Spark now powers an updated version of Meta AI, which users can access online at meta.ai or in the Meta AI app.

“Muse Spark is the first step on our scaling ladder and the first product of a ground-up overhaul of our AI efforts,” a Meta announcement stated.

Spark is designed to be particularly capable in areas tied to everyday personal use — tasks like visual understanding, health, shopping, and social content. Looking ahead, Zuckerberg said Meta is building products that go beyond answering questions, toward AI that acts as agents “that do things for you.”

Future AI models in the Muse lineup will also include new open-source releases.

Muse Spark is the first big product from Meta Superintelligence Labs

The announcement marks the public debut of work that has been underway — and at times turbulent — since last summer. When Zuckerberg first laid out his vision for “personal superintelligence” in a July 2025 manifesto, the ambition was an AI that helps people pursue their own goals rather than one controlled from the top down.

To build it, Meta went on one of the most aggressive hiring sprees in recent memory, personally recruiting more than 50 researchers from rivals including OpenAI, Anthropic, and Google, and bringing in former Scale AI chief Alexandr Wang to lead its new superintelligence research group.

Then, just as quickly, Meta froze hiring altogether — citing routine budget planning — and restructured the team into four smaller units focused on research, superintelligence development, products, and infrastructure. Zuckerberg explained the pivot by saying he believes breakthrough AI work is best done by compact teams who can hold the full picture in their heads, rather than sprawling organizations.

The whiplash raised eyebrows amid broader market jitters about whether the AI boom is sustainable. An MIT study circulating at the time found the vast majority of companies deploying AI were seeing no financial return.

In his original manifesto, Zuckerberg drew a sharp philosophical line between Meta and its competitors, arguing that some AI labs want to concentrate superintelligence and pipe its output to humanity like a utility. Meta sees it differently, he said.

In Wednesday’s Muse Spark announcement post, he once again framed the lab’s founding goal as “putting personal superintelligence in everyone’s hands” — with the underlying belief that empowering individuals, not centralizing intelligence, is how humanity moves forward.

Wednesday’s Muse announcement will be the first concrete product to emerge from these multi-billion-dollar investments. (Meta allocated $72 billion in AI development in 2025 and is expected to spend up to $135 billion in 2026.)

Muse Spark: Benchmark performance

So far, Meta’s Llama family of AI models has lagged far behind its rivals on AI leaderboards. Whether Spark lives up to the superintelligence branding remains to be seen, but after months of hiring drama, restructuring, and big-picture theorizing, Meta has finally put something on the table.

As Zuckerberg put it: “I’m looking forward to sharing more soon.”

As part of its Muse Spark announcement, Meta Superintelligence Labs released its scores on popular AI benchmark tests such as Humanity’s Last Exam (HLE), ARC AGI 2, and GPQA Diamond. These scores could not be independently verified at this time, but Meta did release information on its testing methodology for Muse Spark.

Overall, Meta reported mixed results when comparing Muse Spark to frontier models such as Claude Opus 4.6 Max, Gemini 3.1 Pro High, GPT 5.4 Xhigh, and Grok 4.2, with Muse Spark outperforming on some benchmarks and underperforming on others.

Meta released a table comparing benchmark performance for Muse Spark.

How to try Muse Spark from Meta

Muse Spark is available online now. Desktop users can access the new AI model online at meta.ai. Mobile users can also try Muse Spark in the Meta AI app. Additionally, Meta said that select users will be able to access a private API preview.

To compete with reasoning models from OpenAI, Anthropic, and Google, Meta is also releasing a “Contemplating” mode for Muse Spark, “which orchestrates multiple agents that reason in parallel.”

“This allows Muse Spark to compete with the extreme reasoning modes of frontier models such as Gemini Deep Think and GPT Pro. Contemplating mode provides significant capability improvements in challenging tasks, achieving 58% in Humanity’s Last Exam and 38% in FrontierScience Research.”

Contemplating mode is not yet available; Meta said it will be released gradually at meta.ai, but did not provide a timeline for its release.