Meta’s Big AI Comeback? New ‘Muse Spark’ Model Signals Zuckerberg’s Next Move

Fortune, CNBC, Meta, the New York Times, Business Insider, Wired, the Financial Times, and Bloomberg.

Meta is back in the AI spotlight – or at least trying to be.

The company just rolled out Muse Spark, its first major model from the newly formed Superintelligence Labs, the unit built after CEO Mark Zuckerberg went all-in on AI and spent billions assembling a dream team led by Alexandr Wang. This is the clearest sign yet of what that money bought.

And the message is simple: Meta wants back in the race.

Muse Spark isn’t blowing past rivals like OpenAI, Anthropic, or Google. But it’s holding its own across a wide range of benchmarks – a noticeable step up from Llama 4, last year’s much-criticized release that failed to impress developers or industry watchers.

Still, there’s a catch. Actually, a few.

For starters, Muse Spark isn’t widely available. Unlike Meta’s earlier “open-weight” models that developers could download and tweak, this one is mostly locked inside Meta’s own ecosystem. It’s already powering the Meta AI assistant on its app and website, with plans to expand into WhatsApp, Instagram, Facebook, Messenger, and even its AI-powered Ray-Ban glasses. External access? Limited to a small group of partners for now.

That shift alone says a lot about how Meta’s strategy is changing.

Under the hood, Muse Spark marks a technical pivot. It’s Meta’s first “reasoning” model – meaning it can work through problems step by step instead of just spitting out instant answers. It’s also multimodal, so it can handle both text and images, and even coordinate multiple AI agents working in parallel on complex tasks.

There’s a “thinking mode,” too. Flip it on, and the model essentially breaks problems into pieces, assigns them to different sub-agents, and stitches together a final answer. Meta says this helps it compete with the more advanced reasoning systems from its rivals.

Benchmarks paint a mixed picture. On some tests – like high-level science reasoning – it trails top competitors. On others, especially in health-related tasks, it actually comes out ahead. Competitive, but not dominant.

And yes, skepticism lingers.

Meta has been called out before for presenting benchmark results that didn’t match real-world performance. With Llama 4, the company admitted it used specially fine-tuned versions of the model to boost scores – versions that weren’t available to users. So while Muse Spark looks solid on paper, independent testing will matter.

Behind the scenes, this launch is the product of a major reset.

After the Llama stumble, Meta tore down and rebuilt its AI stack – new architecture, new training methods, new data pipelines. The company claims it can now get similar performance with far less computing power, a big deal given how expensive AI has become.

Zuckerberg didn’t stop there. He poured over $14 billion into Scale AI, brought in Wang as chief AI officer, and kicked off an aggressive hiring spree that reportedly included massive pay packages to lure talent from competitors. At the same time, Meta is spending heavily on infrastructure, with AI-related capital expenditures expected to hit well over $100 billion this year.

There’s also a broader strategy taking shape. While Wang focuses on long-term “superintelligence,” another internal team is working on more immediate, product-driven AI features – a sign Meta is hedging its bets.

Muse Spark sits right in the middle of that.

It’s already being used to power features like smarter search, visual understanding (think snapping a photo and getting instant analysis), coding tools, and even a shopping mode that pulls inspiration from content across Meta’s apps. The idea is to blend AI directly into how people already use Facebook, Instagram, and the rest.

Safety is part of the pitch, too. Meta says the model was rigorously tested and shows strong resistance to harmful requests, including those related to bioweapons. But there’s an interesting wrinkle: external researchers found the model is unusually aware of being tested, sometimes recognizing evaluation setups as “traps.” Meta says it’s not a dealbreaker – but it’s something they’re watching.

For now, Muse Spark feels like a foundation rather than a finished product.

Meta is calling it the first step in a larger “scaling ladder,” with bigger and more powerful models already in development. Whether this one is enough to shift momentum in its favor is another question.

The AI race hasn’t slowed down. Meta just jumped back in.