MinishLab's Blink: 8M-param static embeddings top MTEB at 55.35, run on 31MB CPU weights

tomaarsen · x · 2026-11-18

MinishLab's Blink static embedding thread in one post: blink-8M (English) and blink-multilingual-64M (100+ languages) output 256-dim vectors, MIT-licensed, and drop into Sentence Transformers. blink-8M scores 55.35 on MTEB(eng, v2) and 46.70 nDCG@10 on decontaminated BEIR, leading seven static models at 256d; the multilingual model averages 51.60 on MMTEB's 131 tasks, beating potion-128M (47.67) with half the parameters.

Mechanism is token lookup + average + normalize — no transformer forward pass — with 31MB/256MB float32 weights running on CPU. Speed tests: model2vec 3.6-3.8x faster for single queries, Sentence Transformers 2.2x faster at batchsize=1024; lightweight NumPy-only deployment works without PyTorch.

Related event: MinishLab Releases Blink Static Embeddings for Fast CPU Retrieval(5 posts)→

Original post →

More from Models

Models channel →