Specialized vocab cuts SLM inference cost 77%

huggingface researchers bernardo frança de aguiar, gabriel pimenta de freitas cardoso, and francisco de almeida rocha alves discovered that models specialized for a domain or task by fine-tuning general open-source models like qwen and gemma inherit a vocabulary built for many domains and general applications. they developed a method called vocabx to specialize the vocabulary around the cost of each token rather than its frequency. in an essay grader built on qwen3-4b, they cut measured inference cost by 77% without changing the base model. the researchers found that the tokenizer is usually treated as a fixed property of the base model, but it can be one of the largest cost levers for a model specialized in a single domain.