10 min
How language models work (without the mystique)
Tokens, context windows, and why “the model said” is not the same as “the model retrieved.”
Text in, next-token out
A large language model reads your prompt as tokens (pieces of words) and predicts likely next tokens. Chat products add a system prompt, safety layers, and sometimes tools. The core loop is still prediction.
If you need a fact from last Tuesday’s invoice, the model only has it if you pasted it, a tool fetched it, or it happened to be in training data. There is no built-in filing cabinet.
Context window
The context window is the working memory of one request: the system prompt, your messages, retrieved snippets, and the reply so far. When you overflow it, earlier turns drop or get compressed. Long chats are not infinite memory.
Weights you can download vs APIs you call
Some models are **hosted**: you send a prompt to a company’s API. Some are **open-weight**: you can download files (license permitting) and run them with a local runtime. Hosted is easier. Local keeps the prompt on your machine. Neither is automatically “smarter.”
Model cards and licenses live on the publisher’s site or Hub page. This catalog links there instead of restating licenses.
Related AI sites in the catalog
- Hugging Face — A public hub for model weights, datasets, and demos. Also a place to run and share models without building hosting from scratch.
- OpenRouter — A router that exposes many hosted models through one API. Listed prices on this site come from OpenRouter, not from each vendor’s own price page.
- Ollama — A local runtime for running supported models on your own machine. Weights stay on the computer that runs them.