Meta built its AI credibility on open-source. Now it's walking away from the strategy that made it relevant.
On Wednesday, Meta unveiled Muse Spark, the first model from its newly formed Superintelligence Lab. The release represents a decisive pivot: after years of positioning Llama as the open-source counterweight to proprietary AI giants, Meta is launching its flagship model as a closed product—while signaling future open-source releases under the Muse brand will come, but as secondary offerings, not the main event.
For users, the difference is immediate. Muse Spark taps directly into content across Instagram, Facebook, and Threads, functioning less like a standalone chatbot and more like a personalized intelligence layer woven into Meta's social graph. Ask about a trending location, and it surfaces relevant public posts. Future integrations will weave Reels, photos, and posts directly into answers, with creator attribution. This mirrors how xAI's Grok operates within X—except Meta's social ecosystem is orders of magnitude larger.
The timing is deliberate. Llama never achieved the ranking performance Meta executives hoped for, despite genuine developer enthusiasm. Independent benchmarks consistently placed Llama models below comparable offerings from OpenAI and Google. By launching Muse as a proprietary line, Meta sidesteps the constraints of open-source development—no forced transparency, no community-driven iteration cycles, no reputational risk when the open model underperforms.
Zuckerberg framed it as evolution, not abandonment. In a Threads post, he stated that the Muse family will include open-source releases in the future. But the hierarchy is clear: closed models ship first, open models follow when Meta decides they're ready. This separates Meta from true open-source advocates like Mistral and positions it alongside xAI and Anthropic—companies that release weights on their own terms.
The Superintelligence Lab itself signals ambition beyond incremental improvement. Formed roughly a year ago with the explicit mandate of "delivering personal superintelligence for everyone," the lab operates under different constraints than Meta's foundational AI research. Muse Spark is the first public proof that the lab is building toward AGI timelines, not academic publish-or-perish cycles.
Whether users benefit from Meta's pivot depends entirely on what Muse Spark can do that existing models cannot. The social graph integration is novel but not revolutionary—current functionality surfaces posts related to queries, a feature most users can replicate with existing search. The real test comes with future capabilities: if Muse models can reason across a user's entire social history, synthesize cross-platform behavior, and deliver genuinely personalized intelligence, the closed approach makes strategic sense. If not, Meta will have abandoned its open-source community for a proprietary product that delivers the same generic outputs as competitors.
The $50 billion plus Meta has committed to AI infrastructure suggests this isn't an experiment. Muse Spark is the opening move in a decades-long game. The question is whether users will follow Meta into a closed ecosystem, or whether the Llama community's loyalty transfers to whatever open models eventually emerge from this new lab.