Mark Zuckerberg wants to give away the most powerful thing his company owns. On Monday, Meta published a 6,500-word essay and, alongside it, opened the weights of Muse Spark 1.2, its strongest AI model, meaning anyone can download and run it. It also launched Muse Glimmer, a new family of open models small enough to work on a laptop. Weights are the internal settings that determine how a model behaves, and handing them over is the difference between renting a tool and owning a copy of it.
The argument wrapped around the release is sweeping. Zuckerberg warned against concentrating advanced AI, or "superintelligence," in the hands of a few companies or governments, and cast Meta's open approach as the pro-individual alternative. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," he wrote, sketching a future in which everyone carries an "exceptionally capable personal agent." His daughter, he offered, already codes her ideas and makes videos with it.
The essay reads in places as a broadside against Meta's rivals. Without naming them, Zuckerberg took aim at labs that "only build walled gardens" and at the industry's darker mood: "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future." The targets were not hard to identify. OpenAI's Sam Altman and Anthropic's Dario Amodei have both warned about AI's effect on jobs, and both companies keep their most capable models closed.
Underneath the philosophy sits a sharper competitive calculation, and its name is China. Meta has long championed open models but has struggled to match closed rivals on raw capability. Meanwhile the open-weight field has increasingly been claimed by Chinese labs including Alibaba, DeepSeek and Moonshot. "If Western tech giants only build walled gardens, developers and enterprise builders will naturally pivot to Chinese open-weight models," Counterpoint Research's Neil Shah told CNBC, adding that Meta is well placed to fill that gap. Zuckerberg pressed Washington to help, urging it to cut "friction" on training data and to rethink its stance on distillation, the practice of training one model on another's outputs. That is a pointed request: US officials have moved to crack down on Chinese firms using distillation, and Anthropic accused DeepSeek of exactly that earlier this year. "You can learn from anything you can observe," Zuckerberg countered.
There is a business case running quietly beneath the manifesto. Muse Glimmer's ability to run on a laptop or phone sidesteps expensive cloud data centers, which Shah framed as a way to "outcompete Google, Microsoft and others on the end-user's device." Meta needs the win. Its shares are down roughly 10 percent this year as investors question capital spending that could reach 145 billion dollars, and Zuckerberg is under pressure to show his year-old Superintelligence Labs is delivering.
Critics were not soothed. Anthony Aguirre of the Future of Life Institute asked "what on Earth makes Meta think they, or anyone, could control something exponentially smarter, faster, and more capable?" Even open-source allies hedged. Matt Lane of Fight for the Future called Zuckerberg "annoyingly out of touch on his best day" while agreeing that open AI matters, then warned that only Meta truly wins if everyone runs Meta's models, however open they are.
That is the tension the manifesto never quite resolves. "Open" is offered as a principle, but it also happens to be the one move that converts Meta's weakness in the model race into strength. When the license is free, the advantage shifts to whoever owns the ecosystem built on top. Zuckerberg is betting that can be him.