Why a lab gives away a model

The question worth asking first

Training a model at the top of the open ecosystem costs somewhere between a few million and a few hundred million dollars in compute alone, before salaries. Then the lab publishes the result for anyone to download. If you cannot explain why, you will be repeatedly surprised by what happens next: a licence that tightens, a family that goes quiet, a flagship that stays behind an API while only the small siblings ship.

The releases are not charity and they are not an accident. Each family has a business reason, and the reason predicts the behaviour.

Commoditise the thing next to what you sell

The oldest strategy in software is to give away the complement of your product. If you sell electricity, cheap kettles help you. Meta does not sell model access; it sells attention and advertising, and it needs models to be cheap, ubiquitous and not controlled by Google or OpenAI. Releasing Llama makes the model layer a commodity and keeps the layer above it (the products) competitive. That is a coherent commercial decision, and it explains why Meta could release a strong model for years without ever charging for one.

Alibaba’s Qwen releases sit next to a cloud business. A developer who builds on Qwen and then needs more machines than a laptop has an obvious place to rent them. Google’s Gemma works the same way, with the additional defensive purpose of keeping developers inside a Google-shaped toolchain rather than a Meta-shaped one.

Reputation, recruitment and standard-setting

DeepSeek published unusually detailed papers alongside its weights. The direct return on that is not licence revenue; it is being taken seriously, hiring the people who read the papers, and setting the reference implementation that everyone else compares against. Mistral’s early Apache releases did the same job for a young European company that needed to be believed in before it could sell anything.

Microsoft’s Phi line is closer to a research demonstration: the argument is that carefully built data beats raw scale at small sizes, and the model is the evidence. AI2’s OLMo and EleutherAI’s Pythia are public-good science, funded to be studied rather than sold.

What this predicts

Three things follow, and each has bitten someone.

Openness tracks distance from the frontier. Labs release what is no longer their crown jewel, or what is strategically useful to commoditise. Where a model is the product, it stays behind an API. Expect the newest and largest thing a commercial lab has to be the least open thing it has.

Terms follow strategy, and strategy changes. Mistral shipped its main line under Apache 2.0 and then shipped Codestral under a licence that forbids production use. Llama 4’s terms carried a restriction aimed at entities domiciled in the EU. Stability changed its licensing more than once. None of this was dishonest; the strategy moved and the licence moved with it.

The next version is not promised to you. There is no obligation on any lab to release the successor of a model you built on, under the same terms or at all.

The practical consequences

Do three things and this stops being a risk.

• Keep the weights you depend on. A copy on your own storage is the whole point of open weights. A model that is withdrawn from the hub still runs from your disk. Budget the storage; a 16 GB archive is cheaper than a migration.
• Record the terms you accepted, with the date. Licences and acceptable-use policies get edited. What binds you is generally the version you accepted, and you cannot prove that later without a copy.
• Do not design around a single family. If swapping from a Qwen to a Llama of the same size means rewriting your prompts, your chat template handling and your serving stack, you have built the lock-in that open weights were supposed to prevent.

Renting from a hosted provider is not the opposite of this. The freedom that matters is that you could run the weights yourself and nobody can take the model away from you. Whether you exercise that freedom today is an operational question, not a philosophical one.

The one fragile case

Watch for models released by a company whose only product is that model. The economics have to close somewhere. When it does not, the weights that were free become a paid tier, the licence changes at the next release, or the company is acquired and the new owner has different lawyers. This has happened often enough to be a pattern rather than a warning.