AINEWS
Zuckerberg is releasing Meta's latest models for anyone to download, and openly attacking Google, OpenAI, and Anthropic. The market has already priced the cost.
Meta released the underlying parameters of its Muse Glimmer model on Monday and said the weights for its larger Muse Spark model will follow "in the coming weeks." Both moves mark a return to the open-weights strategy Meta had walked back earlier this year, when it declined to release Spark on safety grounds. The company also published a Mark Zuckerberg essay presenting the strategy as distributing "free" and capable AI to "billions of people," with Google, OpenAI, and Anthropic named as the closed alternative.
The shift has two layers. The first is technical: a researcher or startup can now download Glimmer, run it locally, and modify its behavior. The second is rhetorical, and it is the part that costs Meta nothing. "Most other labs are focused on building AI for companies, governments, or other institutions," Zuckerberg wrote, according to the Financial Times. "So if those labs lead, then the balance of power will favor larger institutions over individuals." The target is named. The message is that the frontier labs are building tools for institutions, while Meta is building tools for people.
What the essay does not say is that the position is a reversal. In July 2025, Zuckerberg told TechCrunch that Meta would "likely" not open source its superintelligence-class models. Earlier in 2026, Meta used a similar safety argument to justify withholding Spark weights. The current commitment to ship those weights within weeks turns the prior stance into a tactical pause. A reader watching the open-vs-closed axis now has to ask which of Meta's stated reasons was the real one.
Meta's share price fell nearly 8% after the company disclosed that AI infrastructure spending pushed free cash flow down 91% in a recent quarter, per the FT. The stock is down roughly a fifth over the past year. Meta still trades near a $1.5tn market capitalization, but the capex bill is no longer abstract. Open-weights releases do not reduce that bill; they cede one lever the closed labs use to monetize compute.
Open distribution does not pay for the data centers. But it does something the closed labs cannot easily copy with a press release: it makes Meta the default choice for any developer or business that wants to run frontier models without a vendor lock-in contract. The CNBC coverage of the Glimmer release describes the model as a move toward "local AI agents," software that runs on a user's own hardware, not on Meta's servers. Each local installation is one fewer query sent to a closed lab's API.
Meta is also spending $1bn on a community-host fund for US data centers, pitched as a concession to local backlash over power and water. The dollar amount is small against the capex bill. The fund is less a financial fix than a political one: it seeds local constituencies for the build-out before regulators and county boards tighten the leash.
The story is a distribution bet, not a manifesto. Meta is choosing openness as a way to compete on a lever the closed labs cannot match, and because each open-weight release weakens the moat of selling API access. The rhetoric about "personal superintelligence" and "billions of people" is the political packaging for the same move. The cleanest falsifier is dated: Meta says Spark weights ship in the coming weeks. If they do, the distribution bet starts to compound. If they slip, the essay will read as cover for a quarter that was already priced in.