Mark Zuckerberg spent $14.3 billion to hire the guy who trained everyone else’s AI. Now we get to see if it was worth it.

Meta dropped Muse Spark this week — the first model from Superintelligence Labs, led by former Scale AI CEO Alexandr Wang. It’s proprietary. It’s consumer-focused. And it represents the most dramatic strategic pivot in Meta’s AI history: the company that championed open-source AI just went closed.

The reaction? Mixed, to put it generously.

The Crisis Behind the Curtain

To understand Muse Spark, you have to understand the failure that birthed it.

Last April, Meta released Llama 4. It was supposed to be a statement. Instead, it was a shrug. Developers weren’t impressed. The open-source community Meta had courted so aggressively moved on to Chinese alternatives and fine-tuned versions of competing models. Zuckerberg was reportedly furious.

Two months later came the most expensive talent acquisition in AI history: $14.3 billion into Scale AI, a 49% stake, and crucially, its co-founder Alexandr Wang. Meta created Superintelligence Labs as an elite internal unit, poached researchers from OpenAI, Anthropic, and Google, and offered some engineers pay packages worth hundreds of millions.

The message was clear: Meta was done playing nice.

Nine months and roughly $125 billion in planned capex later, Muse Spark is the first tangible result.

What Muse Spark Actually Does

Muse Spark is live on meta.ai and the Meta AI app, powering the Meta AI assistant. It’ll soon replace Llama models across WhatsApp, Instagram, Facebook, and Meta’s Ray-Ban smart glasses.

The model is described as “small and fast by design” but capable of reasoning through complex questions in science, math, and health. Some highlights:

  • Visual STEM reasoning — excels at image and video processing tasks
  • Interactive experiences — creating minigames, troubleshooting appliances from photos
  • Health queries — calorie counting from meal photos, health-related Q&A
  • “Contemplating” mode (coming soon) — multiple AI agents working in parallel on complex problems

What Meta didn’t disclose is telling: the model’s parameter count. In an industry where model size is standard measuring tape, that silence suggests either the model is surprisingly small (a flex) or embarrassingly large for its performance (not a flex).

The model also requires a Meta account login — Facebook or Instagram — which immediately raises privacy questions. Positioning Muse Spark as a “personal superintelligence product” while requiring social media login is… a choice.

The Open-Source Betrayal

Let’s call this what it is: Meta abandoned its own playbook.

For years, Zuckerberg positioned Meta as the open-source champion of AI. Llama was the anti-ChatGPT — free, customizable, available to everyone. The strategy made commercial sense: Meta couldn’t charge for AI the way OpenAI or Anthropic could, so it would commoditize models and win on distribution through its 3+ billion users.

Muse Spark flips that script entirely. It’s proprietary. Meta plans to offer paid API access after a “private preview” with select partners. Open-source versions will come “eventually,” but the shift in priorities is unmistakable.

Gartner analyst Arun Chandrasekaran called it a “major shift” that “signals an intention to move away” from the Llama brand. Joseph Ott, CEO of AI startup Samu Legal Technologies, put it more bluntly: “The only reason I would use Llama is that I could fine-tune it.” Without that openness, what’s Meta’s differentiator?

The Real Play: Advertising on Steroids

Here’s where it gets interesting. The chatbot wars are a sideshow. Meta’s real AI play is supercharging advertising.

Muse Spark excels at image and video processing. Meta’s platforms are dominated by visual content. If the model can generate dynamic, personalized ad creatives at scale — tailored to individual users based on behavior across Facebook, Instagram, and WhatsApp — that’s not incremental improvement. That’s a fundamental transformation of digital advertising.

Andrew Boone of Citizens Bank said Meta’s 3 billion monthly users are the “crown jewel,” and the AI opportunity is about making ads more engaging. Morningstar’s Malik Ahmed Khan agreed: “I believe that would be the killer use case from Meta’s perspective.”

Doris Xin, CEO of AI startup Disarray, noted that compared to Claude and Gemini, Muse Spark “definitely feels like it has more of a consumer bent.” That’s not an accident. Meta isn’t building a developer tool. It’s building an advertising engine with a chatbot interface.

The Benchmark Problem

The most damaging early criticism came from François Chollet, creator of Keras and the ARC-AGI benchmark. He accused Muse Spark of being “overoptimized for public benchmark numbers at the detriment of everything else.”

That’s pointed, especially given that Meta reportedly outperforms GPT-5.4 on medical and scientific benchmarks. If those numbers don’t translate to real-world usefulness, they’re marketing, not engineering.

Wang’s response was measured — pointing to positive user feedback on visual coding, writing style, and reasoning, while acknowledging weaknesses on benchmarks like ARC-AGI 2. But the broader issue remains: the AI industry has a benchmark theater problem, and Meta has the resources to push past it. Falling into the same trap would be a massive missed opportunity.

The Endgame

Zuckerberg’s ambitions extend far beyond a chatbot: “We are building products that don’t just answer your questions but act as agents that do things for you.”

That’s the real play — AI agents embedded across Meta’s ecosystem, handling tasks, making purchases, interacting with businesses, all while generating advertising revenue. The chatbot is just the Trojan horse.

But Meta is very late. OpenAI has a massive head start in consumer AI. Anthropic dominates enterprise. Google has distribution advantages that make Meta’s look modest. And the developer ecosystem has little reason to switch to a proprietary Meta offering when free alternatives abound.

The bottom line: Muse Spark isn’t a ChatGPT killer. It’s not trying to be. It’s Meta’s first real attempt to turn AI from a cost center into a revenue engine — and the $14 billion question is whether a company built on social media can reinvent itself as an AI company before the window closes.