Meta is again selling AI freedom: powerful agents on ordinary computers, cheap access for billions and less authority concentrated in a handful of institutions. Yet its return to open weights arrives with a conspicuous boundary around the company’s strongest technology.

The reset began Monday with the release of Muse Glimmer, a 30-billion-parameter model that developers can download, modify and run locally on a Mac or PC with a single consumer GPU. Meta has positioned the Apache 2.0-licensed model for always-on agents that handle multistep work, including code, files and screenshots.¹ Zuckerberg declared that Meta was opening Glimmer’s weights and would soon do the same for Muse Spark 1.2.Mark Zuckerberg (finkd)²

That launch was paired with Zuckerberg’s 6,500-word manifesto, which frames centralization—not runaway AI—as the central danger. “I’m personally more worried about centralization than I am about any of the specific risks others are talking about,” he said, arguing that broad, affordable access would preserve individual agency and American leadership.³ He also pressed for looser restrictions on model distillation and data use, while backing early government access to training information for security reviews.

The pitch has won support from open-model advocates. Yann LeCun amplified praise calling the essay a “carefully considered” and constructive approach to increasingly powerful systems.Yann LeCun (ylecun)⁴ Other backers see local models as a way to reduce dependence on the cloud giants and let businesses tailor AI to their own needs.

But the timeline exposes the tension. Muse Spark debuted as Meta’s closed, frontier-class model in April; Spark 1.1 brought a paid service in July; Spark 1.2 arrived August 5. Glimmer is distilled from that larger system—and, while Meta says it will open weights for a version of Spark 1.2, the more capable Spark line remains the company’s controlled asset.⁵

That distinction matters. Glimmer could process sensitive personal data on-device, making Zuckerberg’s privacy and personal-agent vision more plausible. But “access isn’t the same as ownership”: Meta is still deciding which intelligence users may possess and which stays behind its gate.⁶ Critics also note that open weights are not necessarily fully open source, and that Meta’s rhetoric of individual choice jars with restrictions on rival AI assistants using WhatsApp.

For Meta, the new strategy is both ideological counterattack and competitive repositioning: a US open-weight alternative to Chinese challengers and to the closed-model dominance of OpenAI and Anthropic. Its promise is decentralization. Its product hierarchy suggests decentralization, up to a point.