Context Window
Domain: Models & Providers
Motivation
The context window is the token capacity a model’s input can hold, resolved from the exact model via ctx.llm.resolveModelInfo().context. On the DeepSeek route it defaults to defaultContextWindow (1,000,000) for models without an exact value, and the token meter compares measured pressure against it. The window governs how much history, tools, and prompt fit per turn.
Concrete Example
deepseek-flash and deepseek-v4-pro each carry a 1,000,000-token context window; a request overflowing it fails with CONTEXT_WINDOW_EXCEEDED.
Analogy
The size of the desk you can lay out — only so many papers fit at once.