Dharma-AI has published research on vocabulary specialization for small language models, demonstrating that customizing a model's tokenizer for specific domains can reduce inference costs by 77% without altering the underlying model architecture. The technique, called VocabX, was validated on a Portuguese essay grading system built on Qwen3-4B, where the team replaced the general-purpose vocabulary with one tailored to the grading task, achieving dramatic efficiency gains in production. The key insight is that general-purpose tokenizers create two types of waste in domain-specialized applications: fragmentation, where repeated domain-specific sequences like JSON templates are split across multiple tokens when they could be single tokens, and excess, where the vocabulary contains tens of thousands of irrelevant tokens that still require scoring during inference. Unlike simple token-count reduction, vocabulary specialization directly optimizes the most expensive phase of inference—the decode phase, where generating each token costs 7.4 times more than reading tokens due to memory bandwidth constraints. The optimization proves particularly impactful for small models, where vocabulary parameters account for nearly 10% of total model size. In the decode phase, the model's output layer must read the entire vocabulary matrix to score every possible next token, making unused vocabulary entries a direct memory bandwidth cost. VocabX builds specialized vocabularies around the cost of each token rather than frequency, allowing production systems to lower inference latency and costs without model retraining or architectural changes.