The maximum number of distinct tokens a model can recognize, which must be allocated across general and domain-specific patterns.