1 million tokens. That's DeepSeek V4's stated context window - roughly 750,000 words, or the equivalent of about seven full-length novels fed into a single conversation. The number is impressive on paper. Whether it holds up on actual production code is a different question.
A developer ran a structured test across three real codebases: a 45k-token microservice, a 180k-token monorepo backend, and a 520k-token full-stack application. The tasks weren't toy problems - dependency tracing, cross-file refactors, and bug isolation: the kind of work developers actually reach for a large context window to solve.
Under 150k Tokens: Genuinely Useful
Below the 150k-token threshold, the results were solid. At 45k tokens, the model traced function calls across 8 separate files and reconstructed accurate dependency paths. That's not trivial - multi-file call tracing requires the model to hold a mental map of where functions are defined, where they're called, and what state passes between them. Getting this right at 45k tokens puts DeepSeek V4 ahead of models that struggle with even modest codebases.
The 180k-token monorepo is where the test starts to stress the limits. Context windows degrade non-linearly: a model that handles 45k tokens well doesn't automatically handle 180k tokens four times as well. As context grows, models tend to lose track of information from the early parts of the input - a phenomenon sometimes called "lost in the middle," where recall of content near the beginning and end of a long prompt is strong, but anything in the interior gets muddier.
The 520k Gap Between Claim and Reality
The 520k-token full-stack test represents the real gap between marketing specs and practical behavior. Most models advertised with large context windows were trained and fine-tuned (adjusted using additional data to improve specific behaviors) primarily on shorter inputs. The claimed maximum is often a technical ceiling, not a performance guarantee. Degradation at these lengths typically shows up as confident but wrong answers - the model sounds precise about cross-file relationships that it has actually confused or dropped.
For developers evaluating DeepSeek V4 for code-level work, the data suggests treating the 150k range as the practical ceiling for tasks that require accurate recall across the full codebase - not the advertised 1M. That's still a genuinely large working set, covering most medium-sized projects end to end. But it's a meaningful gap from the headline number, and one worth accounting for before routing large monorepos through the model on production tasks.
DeepSeek hasn't been unusual here. Context window claims across the industry routinely outpace real-world recall performance. The more useful question to ask of any model isn't "how large is the context window?" but "at what size does answer quality start degrading on tasks like mine?" This test begins to answer that for DeepSeek V4.