Specializing SLM vocabularies can cut inference cost by 77%

A preview of the VocabX paper shows that rebuilding the Qwen3-4B vocabulary for a single domain and output format can greatly cut inference cost. In a Portuguese essay-grading system, the method designed the vocabulary around token cost rather than frequency; with the same base model, data and training recipe, measured inference cost fell 77%, while shrinking the roughly 151,669-entry vocabulary scored at each generated token.

• Ornith-1.5-122B released: self-improving model built on Qwen2026-10-20
• U-Space offers label-free uncertainty detection for LLMs2026-10-12
• FairStress: New benchmark evaluates fairness in LLM decision-making2026-10-20
• Researchers show edited open models can hide backdoors2026-10-20
• Retrained Qwen3-4B removes AI writing tics2026-10-19