I’m comparing DeepSeek distilled reasoning models for a local setup and keep finding memory estimates that don’t say what context length they assume. A model that loads successfully at a short context may need much more memory once the conversation gets longer.

For people running these models, what GPU memory have you needed for different quantizations at a context length you’d actually use? I’m trying to separate the model’s basic weight footprint from the extra memory required by the KV cache.