I’m testing prompts where the relevant details are spread across several sections of a long document, with similar names and a few exceptions mixed in. How well does DeepSeek keep track of those details across a long context window?

For people using it on specifications or collections of notes, does accuracy fall off gradually as the prompt grows, or are there particular kinds of information it tends to lose or combine?