The Door Meta Just Closed

On 8 April, Meta released Muse Spark, the first model out of its Meta Superintelligence Labs. The division is run by Alexandr Wang, formerly of Scale AI, hired last summer in the nine-figure compensation round that made headlines even in an industry inured to them. The model is small, fast, tuned for science, maths, and health reasoning. Meta has told investors it will spend between $115 and $135 billion on AI infrastructure in 2026.¹

The detail that matters, though, is a sentence that appeared almost in passing in the launch materials: Muse Spark is proprietary. Meta says it may open-source future versions. It does not commit to doing so.²

The detail that matters, though, is a sentence that appeared almost in passing in the launch materials: Muse Spark is proprietary.

I want to stay with that sentence, because the ordinary way to cover this release is to treat it as a product launch, another frontier model, another capex number, another entry in the race. That framing is not wrong but it misses what actually happened. For roughly three years, Meta was the counterweight to closed AI. Llama was imperfect, its license was not quite open-source by the formal definition, and plenty of researchers said so at the time. But Llama 2, Llama 3, and the weights Meta released alongside them were the single largest reason that university labs, smaller companies, governments outside the US, and independent researchers had frontier-adjacent capability at all. Mark Zuckerberg made the case for this publicly and repeatedly, arguing that open weights were both strategically sound for Meta and broadly good for the field.³

Muse Spark is the first model from the lab that now represents Meta's most serious AI ambitions, and it is closed. That is the story. The product is secondary.

The frame the coverage took

Most of the launch coverage framed Muse Spark in competitive terms: Meta has a science-and-reasoning model now, how does it stack up against OpenAI's o-series and Anthropic's Claude, is Wang's lab delivering, will the capex pay off. These are reasonable questions. They are also the questions Meta's communications team would like to be asked, because they position the release inside a race where Meta is a participant rather than inside a longer story where Meta was, until this month, playing a different role entirely.

The race frame treats openness as a tactical choice firms make based on where they sit in the market. Leading firms close; trailing firms open, to commoditise the leader's position. On that reading, Meta's shift is simple: Meta now thinks it is closer to the frontier than it was, or wants to be, and so closes. Nothing to see.

I don't think that frame is adequate to what's happening, and I want to explain why.

Who the open-weights ecosystem was actually for

The people who relied on Llama were not, mostly, Meta's competitors. Anthropic and OpenAI did not need Llama; they had their own frontier models. The people who needed Llama were the ones who could not build a frontier model and could not afford frontier API access at the scale their work required.

That includes academic researchers studying model behaviour, bias, safety, and capability, work that requires access to weights, not just to an API, because you cannot probe a model's internals through a rate-limited endpoint. It includes researchers in countries where US API access is restricted or expensive. It includes small companies building products in regulated sectors where sending data to a US hyperscaler is not a legal option. It includes government labs in the UK, France, India, and elsewhere that have built national AI capacity on top of open weights because building from scratch was not feasible. It includes the hobbyist and independent research community that has produced, over the last two years, a meaningful share of the field's public understanding of how these models actually work.

That ecosystem was not a gift. Meta had strategic reasons for building it. Open weights weakened OpenAI's and Google's moats, made Meta attractive to AI researchers who cared about publication and impact, and built the dependency on Meta's training decisions that now shapes a large fraction of downstream AI development worldwide. Everyone who argued for Meta's open-weights strategy inside Meta could point to those returns.

But the ecosystem existed. And the people inside it made decisions, career decisions, company decisions, research decisions, national policy decisions, on the assumption that Meta would keep releasing weights. They were not told this was conditional. They were told, repeatedly and publicly, that openness was Meta's philosophy.

What "may open-source future versions" actually means

The phrase sounds reassuring. It should not. When a company moves from "we release weights" to "we may release weights," the burden of proof has inverted. Under the old arrangement, a closed release from Meta would have required explanation. Under the new arrangement, an open release requires a business case to be made internally, every time, against the default of closure.

I have watched enough corporate language drift to know how this story ends. The next model will be closed because it is too new. The one after that will be closed because of safety concerns. The one after that will be closed because of competitive pressure. At some point, a Llama-branded model will be released with weights, and Meta will point to it as evidence that the philosophy is intact. But the frontier work, the models that matter, the ones the Superintelligence Labs are being built to produce, will be closed. The open releases will become legacy maintenance.

I would like to be wrong about this. The evidence that I am wrong would be a clear, public, written commitment from Meta about which categories of model will be released with weights and under what conditions. Muse Spark's launch materials do not contain such a commitment. They contain an aspiration, phrased to preserve maximum flexibility.

Why this is a consequences piece, not a competition piece

If you read the release as a competitive move, the affected parties are Meta's competitors and Meta's shareholders. Competitors must now respond to a new model. Shareholders must evaluate whether $115–135 billion of 2026 capex will be recouped. This is the frame Meta's investor relations team will use. It is also the frame most of the tech press will adopt, because it is legible and the numbers are big.

If you read it as a shift in the distribution of AI capability, the affected parties are different. They are the research labs that cannot replicate frontier training runs. They are the regulators in jurisdictions that were building oversight regimes around the assumption that weights would be available for audit. They are the smaller firms that were building product on top of Llama and now face a strategic question about whether to bet on future Meta releases or migrate to one of the remaining open-weight providers, of which there are fewer than there were two years ago, and most of which are smaller and less well-capitalised than Meta.⁴

The distributional question is: what does it mean for the global AI research and development ecosystem that the largest open-weights provider is signalling that its most serious work will be closed? I don't think the answer is catastrophic. There are other open-weights efforts, Mistral, the Qwen series from Alibaba, DeepSeek, the German and French national efforts, the UK's AI Research Resource investments. But each of these was, in part, possible because Meta had set a ceiling on how closed the frontier could plausibly become. Remove that ceiling and the economics of open-weight development shift against the people doing it.

There is a particular irony here that I want to name, because it matters. The argument Zuckerberg made for Llama's openness was that concentrated AI capability was dangerous and that widely distributed capability was safer. Whatever you thought of that argument, and I thought parts of it were self-serving and parts were correct, it was an argument against the concentration of power. Meta Superintelligence Labs, with its $100M+ hiring packages and its closed first release, is a bet that concentration, not distribution, is where the returns lie. The argument has not been publicly retracted. It has simply stopped being acted on.

What to watch

Three things, over the next twelve months.

First, whether any Muse-branded model is ever released with weights. The naming matters. "Llama" was the open line; "Muse" appears to be the new frontier line. If the frontier line stays closed and the open releases are confined to older architectures, the de facto policy is clear regardless of what the press releases say.

Second, whether the researchers and labs that built on Llama adjust their public positions. Many of them have been circumspect about Meta's openness, not wanting to jeopardise a relationship they depended on. If they begin saying in public what many have said in private, that the openness was always partial and always revocable, that will tell us something about how the ecosystem is reading the signal.

Third, whether governments respond. The EU AI Act, UK regulatory work, and several national AI strategies assumed a mixed ecosystem in which some frontier capability was auditable because weights were available. A shift toward closed frontier models from the largest open-weights provider is a structural change in what regulators can actually see. I would expect, and would watch for, policy responses that try to require some level of transparency for models above a capability threshold, whether through mandatory model cards, safety case disclosures, or evaluation access regimes.

The launch of Muse Spark is being covered as Meta entering a race. It is at least as accurately described as Meta leaving a position, the position of the large, well-capitalised, strategically open counterweight, that no one else is obviously placed to fill. That vacancy is the story. The model is the occasion.

Footnotes

¶