The 179x Tax on Non-English AI

An EMNLP 2026 paper measured total inference energy across eight languages and found a 179x gap between English and Pashto for identical request sets. The penalty has two compounding causes — per-token energy cost up to 8.3x higher in complex-script languages, and many more tokens needed for low-resource ones — and both directions also correlate with lower task accuracy.

Read more →

The Case for Translation-First Models

Bilibili's Index LLM Team released Index-Translate, a 35B MoE model family for 150 languages under Apache 2.0. Only 3B parameters activate per token. The family includes specialized checkpoints for speech translation, syllable-constrained dubbing, and full-document translation — and the MoE variant matches frontier generalist models on FLORES at a fraction of their inference cost.

Read more →