Language models and open weights

In one paragraph

A language model is software that reads a question, written in ordinary language, and writes an answer in the same way. It gets that ability from being trained on very large amounts of text beforehand, which teaches it the patterns of language well enough to produce a fitting reply to something it was not trained on directly. Some language models are offered as a service you send questions to over a connection; others are released as files you can download and run yourself. That difference matters more than it sounds.

Why it matters

Where a language model runs, and who controls the file it runs from, decides who can see the questions and documents that pass through it. A model offered only as a rented service means every question travels to wherever that service runs. A model released as a downloadable file can be installed inside your own environment, so questions and documents can stay inside it throughout. For an organisation that has to be careful about what leaves its own boundary, that difference is often the deciding factor in how AI gets used at all.

How it works

• A language model is trained once, on a large body of text, to learn the patterns of language and the knowledge embedded in it.
• After training, it answers new questions by predicting a fitting continuation of the text it is given, one small piece at a time.
• Some models are only reachable as a service: you send a question in, and an answer comes back, with the model itself staying inside the service provider's infrastructure throughout.
• Other models are published with their weights, the numbers that define what the model has learned, available to download. This is what people mean by open weights.
• A model you hold as a file is one you can install, run, and keep exactly as it is until you choose to change it, rather than one that can change under you without notice.

What it looks like in practice

The question usually arrives as a choice about documents. Sending those documents to a rented service means trusting that service with material the organisation would rather keep inside its own boundary. Choosing a model available as a file instead means the whole exchange, question, document, and answer, can stay on infrastructure the organisation controls, with the same model running until the team decides it is time to move to a newer one.

How this connects to our work

We generally install models available as files, because that is what makes private deployment possible: the whole exchange stays inside your own environment. How much work a model can get through depends on tokens and how inference is measured. Choosing which kind of model answers which kind of question, and writing that choice down as a policy, is part of private AI setup.