minishlab's Blink: 8M-parameter static embeddings run on CPU without a transformer pass

tomaarsen · x · 2026-11-18

minishlab released the Blink family of static embedding models: English blink-8M (7.68M params, 31MB float32) and blink-multilingual-64M (100+ languages, 256MB).

Instead of a transformer forward pass, they look up token vectors, average them, and normalize—producing 256d embeddings extremely fast on CPU. Both are MIT licensed and load directly via model2vec, model2vec-rs, or Sentence Transformers. The author demoed blink-8M retrieving the right help article.

Original post →

More from Models

Models channel →