Context Window ☆

The context window is how much a model can consider at once. GPT-3 held 2K tokens; Gemini holds a million; long-context serving is why the KV cache (the stored state for every token in context) dominates inference memory and why chips like Rubin CPX exist just for the 'prefill' of huge contexts. Long context competes with retrieval (RAG) as the way to give models knowledge.