Zero. That is the number of people outside Meta who can download, inspect, or run Muse Spark, the company’s most capable AI model. Meanwhile, Glimmer — Meta’s newly released open-weight model — is available to anyone with the hardware to run it. Mark Zuckerberg frames this as AI “for everyone.” As someone who has spent years studying model architectures and the systems built around them, I want to unpack what that framing actually means, and what it conveniently leaves out.
The Two-Tier Strategy, Decoded
Let’s be precise about what happened. Meta shipped Glimmer as an open-weight model. That means the trained parameters are downloadable: you can run inference locally, fine-tune on your own data, and build agents on top of it without asking Meta’s permission. That is genuinely valuable. Open weights are how independent researchers audit behavior, how startups build without platform risk, and how the broader community stress-tests safety claims that companies would otherwise self-certify.
But Muse Spark, the proprietary model, stays inside Meta’s walls. And this is where the “for everyone” rhetoric starts to strain. If your best model is closed and your second-tier model is open, you have not democratized AI. You have democratized a capability floor while retaining a capability ceiling for yourself.
From an architecture standpoint, this matters more than it might appear. The gap between an open model and a closed frontier model is not just benchmark points. It shapes what kinds of agent systems the outside world can build. Agents are only as capable as the reasoning engine underneath them. If the strongest reasoning engine is API-gated and unmodifiable, then the most capable agents will always be built inside — or on the terms of — the company that owns it.
Open Weights Are Not Open Source
A technical distinction the marketing tends to blur: open weights give you the artifact, not the recipe. Without training data, training code, and the full methodology, you cannot reproduce the model, verify what it was trained on, or fully understand its failure modes. You can probe it, fine-tune it, distill it — all useful — but you are working with a black box that happens to be sitting on your own GPU instead of someone else’s.
I do not say this to dismiss Glimmer’s release. Open weights are strictly better for the research community than a closed API. Interpretability work, red-teaming, and agent orchestration research all become tractable when you have direct access to logits and activations. My lab-adjacent colleagues will get real mileage out of this. But “you can run our model” and “AI is for everyone” are different claims, and only one of them is supported by the release itself.
The Strategic Logic Behind Selective Openness
Why would a company open one model and close another? The pattern is not mysterious. Releasing a strong open model commoditizes the layer of the stack where your competitors charge money. If solid open models exist, the market price of mid-tier inference trends toward the cost of compute, which pressures rivals whose business depends on API revenue. Meanwhile, keeping your frontier model closed preserves your differentiation and your ability to monetize the top end.
This is a coherent commercial strategy. It may even be a defensible one. But it is a strategy, not a philosophy. The question posed in the headline — does Zuckerberg really believe AI is for everyone — may be the wrong question. Belief is unfalsifiable. Behavior is not. And the behavior says: some AI is for everyone, and the best AI is for Meta.
What Would Genuine Universal Access Look Like?
If I were grading commitments to accessible AI, I would look for a few things:
- Frontier parity: releasing open models that are the company’s actual best, not a tier below.
- Training transparency: documentation of data and methods sufficient for independent reproduction and audit.
- Stable licensing: terms that do not quietly restrict commercial use or reserve the right to revoke access later.
- Compute pathways: acknowledgment that weights without affordable compute still exclude most of the world.
Glimmer clears none of the first three fully, and the fourth is a structural problem no single release solves.
So credit where due: Meta’s open-weight releases have consistently pushed the field in a more accessible direction, and Glimmer continues that. But “for everyone” is a claim about the whole portfolio, and as long as Muse Spark sits behind a locked door, the honest description is “for everyone, up to a
🕒 Published: