I’m evaluating DeepSeek for questions over long documents and keep wondering whether a fact is easier to retrieve when it appears near the beginning or end of the prompt. A model can accept a large context window, but that alone doesn’t tell me how consistently it uses information buried in the middle.

Has anyone tested this by moving the same key passage to different positions while keeping the question and surrounding text the same? I’d be interested in which DeepSeek model and serving setup you used, and whether you saw a meaningful difference between short and very long contexts.