Meta Llama

Meta Llama is the family of large language models released by Meta with openly downloadable weights, so that anyone can run, fine-tune or build a product on the model rather than call it through one vendor’s service. It matters here because assistants built on it can be asked about your business, and you may not know which ones exist.

What open weights change

A model whose weights are published runs wherever someone deploys it. There is no single assistant to probe, no single crawler to allow or block, and no single place to check what it says. One deployment may search the web and fetch pages before answering; another may write from training data alone; a third may be fine-tuned on a company’s own documents. By contrast, an assistant run by one vendor documents its behavior in one place: OpenAI’s crawler documentation names the product tokens ChatGPT uses to index, to train and to fetch on a user’s behalf. A self-hosted Llama deployment that fetches pages identifies itself however its operator configured it, if at all.

For a business owner the practical consequence is that the same page has to serve readers you cannot enumerate. The only lever that reaches all of them is the page itself: reachable without a challenge, with a passage that answers the question and an identity a machine can pin down.

What this site has and has not measured

This site’s published probe used Claude, and its recurring measurement puts its question panel to Claude as well, as the methodology page sets out. It has not measured any Llama-based assistant, and nothing on this page should be read as a rate for one. The mechanism above follows from what open weights are; the numbers, when they exist, will be reported as what they are.

A common question

How should a business prepare for assistants built on openly downloadable models?

By fixing the page, since nothing else reaches all of them. Published weights run wherever someone deploys them, so there is no single assistant to probe, no single crawler to allow or block and no one place to check what is being said; one deployment searches the web before answering, another writes from training data, a third is tuned on a company's own documents. The page has to be reachable without a challenge, carry a passage that answers the question, and pin down an identity a machine can hold on to.

Sources

Vendor crawler documentation, for the contrast: OpenAI’s crawler documentation. This entry carries no figures and this site has not measured a Llama-based assistant.