Mark Zuckerberg spent 6,500 words on Monday arguing that the greatest danger in artificial intelligence is not a model going rogue but a handful of companies owning the thing outright. “The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he wrote, in an essay that names no rivals and is unmistakably about two of them.
The argument is coherent. It is also, conveniently, the argument that a company running third in the frontier race would want everyone to accept.
Read the Release Schedule, Not the Philosophy
The essay did not arrive alone. Meta shipped Muse Glimmer the same day, a 30-billion-parameter multimodal model tuned for agentic tool use and coding, released under an Apache 2.0 license that permits commercial use and downstream training with no strings. Quantized to four bits it fits under 20GB, which means it runs on a single consumer graphics card. Meta also said it intends to publish weights for a version of the more capable Muse Spark in the coming weeks.
That sequencing matters more than any paragraph in the essay, because Meta spent the spring going the other direction. Llama was open. Muse Spark, which succeeded it in April, was not. The company that built its AI reputation on open weights closed the door, watched what happened, and has now reopened it with a manifesto attached.
Nothing in the intervening months made concentration more dangerous. What changed is Meta’s position. We covered the frontier capability gap Meta has been trying to close earlier this summer, and it did not close. When you cannot win on capability, you change what the competition is about.
Commoditize the Complement
The strategy here is old enough to have a name, and it has nothing to do with safety.
OpenAI and Anthropic sell inference. Their revenue is a function of tokens processed at a price they set, defended by the fact that the best models are expensive to run and live behind an API. That is a real moat as long as the gap between the best model and the best free model stays wide enough to justify the invoice.
Meta does not sell inference. Meta sells advertising, and every dollar an enterprise sends to Anthropic is a dollar that does not become Meta infrastructure, Meta distribution, or a Meta-shaped developer ecosystem. Releasing a genuinely useful model for free does not cost Meta a revenue line, because it never had that revenue line. What it does is push the floor price of a large category of AI work down toward the cost of the electricity to run it.
If a 30B open-weight model handles your agentic tool calls, your coding assistant and your evaluation harness competently on hardware you already own, a meaningful slice of paid API demand simply stops existing. Not the frontier slice. The boring, high-volume, margin-generating middle.
Zuckerberg is not wrong that distributed capability is safer than concentrated capability. He is also, separately, running the single most effective commercial play available to a company that is behind: make the thing your competitors charge for into something people expect for nothing. Both statements are true. Only one of them is in the essay.
The Distillation Tell
The clearest evidence that this is a competitive document aimed squarely at OpenAI and Anthropic sits in the section about distillation, the practice of querying a rival model to extract what it knows and train on the answers. Zuckerberg defends it directly: “I think it is important to protect the principle that you can learn from anything you can observe.”
Consider who needs that principle protected. A lab with the best model in the world wants distillation treated as theft, because its capability lead is the asset and distillation is the mechanism by which the lead leaks. A lab in second or third place wants distillation treated as learning, because it is the cheapest path to closing a gap it cannot close with compute alone. Meta has enormous compute. What it has not had is the lead.
Framing that as an open-inquiry principle rather than a competitive position is good rhetoric. It is not a neutral observation, and the terms-of-service fights it anticipates are going to be expensive for somebody.
The rest of the essay is more conventional. Zuckerberg described a labor market of “a larger number of companies with fewer people,” each person working alongside what he called a personalized tutor and coach with a PhD in every subject, and committed Meta to free tools for billions of users plus a fully private mode for personal AI agents. Those are positioning statements aimed at regulators and consumers, and they cost nothing to make.
What It Costs to Mean It
The open question is whether Meta can afford this posture through a downturn in enthusiasm.
Open weights are a gift with no invoice attached, funded out of a capital budget that has to answer to shareholders eventually. Meta’s own FY27 capital expenditure guidance put the scale of that spending in front of investors in July, and the honest read is that the advertising business is underwriting an AI program whose direct revenue contribution is hard to point at. Giving the output away makes the accounting cleaner in one sense, since nobody expects a free product to show a margin, and harder in another, because the return has to show up somewhere else and on someone else’s timeline.
Meta abandoned open weights once already this year when the competitive logic pointed the other way. The manifesto is an argument that it will not do that again. The April decision is an argument that it might.
Watch what happens if Muse Spark’s weights slip past the promised window, or arrive with a license carve-out that Apache 2.0 does not have. The essay will still read as principle. The release notes will tell you whether it was.