Why AI labs publish open weights, and what it costs you

Open weights are usually discussed as a moral position. They are a commercial one, and the companies publishing them have been reasonably clear about that if you read what they say rather than what the community hears.

This is not an argument against downloading them. It’s an argument for understanding what you’re being given and why.

Who publishes, and what they keep

The pattern this year is consistent. Labs release a capable model a tier below their best, and keep the best one behind an API.

Meta published a 30B model for consumer hardware while its strongest coding model stayed closed and powered the agent it sells to enterprises.

Give away the tier below your best, sell the product built on your best. That is not generosity. It is a funnel.

Thinking Machines shipped open weights from a standing start, which is a distribution play by a company with no installed base to protect.

Both moves make sense. Neither is charity, and treating them as charity leads people to skip the questions they’d ask any other supplier.

What open weights genuinely give you

The benefits are real and worth stating before the criticism, because they are the reason this matters at all.

A file on your disk cannot be deprecated on somebody else’s schedule. It runs where you run it, which resolves a category of data questions rather than mitigating them, as our guide to running a model locally covers.

The capability argument has also stopped being a compromise. Two of the six labs at the top of the 2026 AI Index ratings publish weights, which our piece on parity at the frontier works through.

Researchers have started testing that in narrow professional domains rather than on general benchmarks, asking directly whether open-weight models compete on financial text comprehension. That the question is worth asking is itself the finding.

The safety argument cuts both ways

The strongest criticism of open weights is that publishing them removes any ability to withdraw a capability, and the strongest defence is the same fact.

Government evaluation bodies now test frontier models, and none of them can stop a launch. With open weights they cannot even stop distribution, since a file that is out is out permanently.

Regulators have leaned toward transparency rather than restriction so far. Four labs agreed a voluntary testing framework at the White House this month, and it doesn’t distinguish between open and closed release at all.

Against that, published weights are the only models outside researchers can actually inspect. Every mechanistic interpretability result on a closed model depends on the lab granting access; on an open one it depends on a download.

Both things are true, and anyone claiming the question is settled in either direction is arguing a position rather than reading the evidence.

What the word open is doing

What you get	What you do not
The weights	The training data
Permission to run it	Always permission to run it commercially
Ability to fine-tune	Any account of what was filtered out
Independence from an API	Independence from the licence terms

The data column is the one with legal consequences. Rulings so far have turned on how training material was obtained, which our piece on the copyright cases sets out, and a permissive licence on weights says nothing about the corpus.

Licence terms are the other trap. Several widely used releases restrict commercial use or cap user numbers, and a term you discover at scale is a term you discover expensively.

The dependency you keep

Running weights yourself removes an API dependency and replaces it with several quieter ones that are easy to miss until they bite.

You now depend on an inference engine, a quantisation format, a hosting provider for the download, and a community that keeps all three working together. None of those is contractual.

You also inherit the operational work. Uptime, patching, capacity planning and someone available when it breaks are now yours, which our guide to the self-hosted stack works through layer by layer.

That trade is often worth making. It is not the absence of a trade, and open weights get discussed as though it were.

The strategic reading

Publishing weights commoditises the layer you are not selling. If your money comes from advertising, or from an enterprise agent, then free capable models damage your competitors more than they damage you.

Meta has been explicit that the goal is personal assistants running on hardware people already own, which CNBC read as strategy rather than benchmark-chasing. Local inference is free to Meta and expensive to anyone selling tokens.

That’s a legitimate strategy and it has genuinely good side effects. It also means the supply of open models depends on that calculation continuing to hold, and nobody’s promising it will.

Which is the actual risk. A community that treats open weights as a movement rather than a business decision has no plan for the year the decision changes.

There is precedent for the reversal. Companies have closed previously open projects when the competitive maths shifted, usually with a licence change rather than an announcement, and the users found out at renewal.

Regulation may also intervene. Marking obligations under the EU AI Act are straightforward for a provider serving an API and considerably harder for one shipping a file that anyone can modify.

What to do about it

Read the licence before you build, and keep a copy of the weights you depend on rather than assuming the download stays available.

Pin the version too. Open-weight families iterate as fast as closed ones, and the file you validated is not necessarily the file at that URL next quarter.

Ask providers what they will state about training data, because that answer is your exposure and it is rarely volunteered.

Test the model on your own tasks before adopting it, since a capable open model and a capable open model for your work are different claims, and only one of them is on the model card.

And watch whether the tier gap narrows. If labs start publishing their best models rather than the tier below, the argument that this is a funnel gets weaker. Nothing in 2026 suggests that yet.