- The VocabX method rebuilds a model’s vocabulary for one domain; the authors report 77% lower cost in one case.
- Results are from the creators, and the paper is still under review.
In short: the Dharma-AI team shows that an AI model serving one narrow task can run noticeably cheaper if its vocabulary is fitted to that task. The numbers are from the authors.
What changed
- The VocabX method builds a vocabulary for a specific domain without changing the base model, the data or the training recipe.
- In one test, grading Portuguese essays on Qwen3-4B, cost fell by 77.0%.
- The average response shrank from 350 to 101 pieces, and the benchmark score rose from 0.703 to 0.810.
- The vocabulary shrank from 151,669 to 64,767 entries.
Test conditions
Seven versions were trained identically, differing only in vocabulary. Each handled 500 requests on an RTX 5090. Across five public domains and three model families (Qwen3, Llama-3.1, Gemma-3), the method beat three reference methods on compression in all 15 combinations.
How to read this
The paper is still under review and the numbers may change. It was published by the company that built the system. The authors note the method does not make the model smarter, and the 7.4x cost ratio between generating and reading comes from one case and depends on hardware. More on running models yourself in the local AI hub.