Mark Zuckerberg Gives Llama Away, and Charges Nothing

I once watched a man give away free coal in a town where he owned every stove. Nobody called it charity for long. When Meta began handing out the weights of its large language models to anyone with a hard drive, the industry reacted with genuine confusion. The prevailing belief was that model weights were the crown jewels, the one asset no firm would let out of the vault. Giving them away looked like a category error, and I have made a few myself.

Zuckerberg’s explanation was blunt. His company had been burned before by platforms it did not control, and he would not let that happen with the next one. I have heard longer speeches that said less.

The bookkeeping behind the generosity

Meta’s business depends on building products cheaply at enormous scale. If the thing a company needs is rented from a competitor — or from a partner one bad quarter away from becoming a competitor — then its costs and its road map belong to somebody else.

Open weights fix that at the root. Even if Meta never runs a community model in production, a free alternative puts a ceiling on what it can be charged. That is not sentiment. That is a fence post driven into the ground, visible from the road.

• Turning the layer below you into a commodity is the oldest platform play in the book.
• Developer goodwill is an asset you can hire with and lobby with.
• It forces rivals to compete on the product rather than on who controls the door.

The objection that keeps its own hours

The objection worth taking seriously is that a released model cannot be recalled. Once the weights are out they can be fine-tuned for purposes their makers cannot prevent, which is rather like letting a horse out of the barn and then publishing a memorandum about the barn door. Meta’s answer is that the marginal risk is small when comparable ability is already rentable through commercial APIs. The argument has a cold arithmetic to it.

Researchers who study misuse dispute the point, and I do not undertake to settle it, having been wrong before and expecting to be again. But the argument has become the working position of the industry, and several other laboratories have adopted variations of it. That is how a contested claim becomes a settled one — by repetition and a busy schedule.

Nobody reads the invoice

The least discussed part of the strategy is the bill. Meta’s capital spending on AI infrastructure has grown to a figure that would have been laughed off the page a few years ago, and the company has had to explain to investors why a free product justifies it. The honest explanations are always the shorter ones.

Zuckerberg’s answer is that the models improve every existing product and spawn new ones besides, and that the return arrives as engagement rather than as licence revenue. So far the market has been willing to accept it.

An asset that belongs to nobody

Releasing weights produces something no purchased model can: developers who have built on your work, written the tutorials, and patched the bugs in the tooling around it. That crowd grows the way a town grows, and no single company owns the town.

It also opens a door into enterprises that would never have taken a vendor’s meeting. Teams that prototype on a free model tend to keep using it, and the organisation that trained it never has to send a salesman up the front walk. I have known companies spend a fortune to acquire what this one gets by leaving the gate open.

The safety question nobody has closed

The claim that open releases are safe because equivalent capability is already for sale assumes that capability is the only thing that matters, which is a large assumption wearing a small coat. Fine-tuning a local model with no oversight, no logging and no terms of service may be a different class of risk from renting an API that logs and moderates. The ability may be the same and the consequences may not be.

There is no good empirical answer to that yet, and the industry has largely gone ahead in the meantime. That is the honest state of the question, and I would rather leave it there than polish it into something it is not.