Context window

What fits inside a context window?

A context window is the token budget a model can use for a request and its generated continuation. It places a practical limit on how much material can be considered together. Instructions, messages, documents and tool results all compete for that space, and the application must also leave room for the answer it needs.

The budget is measured in tokens rather than pages or characters. A page count does not translate into a dependable token count because document formatting, language and tokenization differ. Use the relevant tokenizer or measured API usage when assessing a real workload.

Model capability and served limits

A model card may describe an architectural context capability, while a particular service exposes a smaller configured limit. These are separate facts. Check the limit for the endpoint you will call rather than assuming that a number in a model announcement is also a service guarantee.

IndicStack's live Qwen3.8 27B model has a confirmed 262K context. Unknown served limits for other catalog entries are omitted from public fields. A long-context capability filter describes a model category; it should not replace verification of the specific service configuration.

A large window also does not guarantee that a model will use all supplied information accurately. Evaluate questions whose answers appear in different parts of the material. Check whether irrelevant passages distract the model and whether the answer cites the evidence your application expects.

How should an application manage context?

Reserve an output budget before filling the request with source material. Keep essential instructions clear and select conversation history deliberately. When a tool returns a very large result, consider whether the next model step needs the entire result or only a relevant subset.

For document workflows, retrieval can select useful passages instead of adding a whole collection to a prompt. Chunking can make long material manageable, but the application still needs to preserve relationships across boundaries. A useful answer may depend on information from several sections.

Test requests near your intended operating size and handle oversized inputs explicitly. Let the product explain what the user can change, such as narrowing a document selection. Silent truncation can remove the very evidence needed for an answer, making failures harder to diagnose.

Related terms