Context length

Also written: context, context window, n_ctx, -c, num_ctx

The number of tokens (prompt plus reply) a model can hold at once. Longer contexts need a bigger KV cache, so local tools start with short ones like 4,096 tokens.