Live page · Day archive

Embeddings Are Overpriced? A Frozen 16 MB Vocabulary for Small and Ternary Language Models

Read the original at arXiv

Summary

Researchers replace the large embedding table in Gemma 3 270M with a frozen 16 MB vocabulary, achieving comparable performance and faster inference on laptop CPUs.

Carried by: arXiv, HF Daily Papers. First seen: 2026-10-19T05:30:03Z.