Blink CPU speed test: model2vec 3.8x faster solo, Sentence Transformers 2.2x faster at batch 1024

tomaarsen · x · 2026-11-18

tomaarsen benchmarked Blink static models on an i7-13700K with short texts: for single queries, model2vec was 3.6-3.8x faster across both models; with batchsize=1024 for both, Sentence Transformers was 2.2x faster. He expects the Rust model2vec-rs implementation to beat both.

For lightweight deployments, model2vec runs the same models via NumPy without PyTorch, and since Blink embeddings are normalized, a dot product yields cosine similarity.

Original post →

More from Infra

Infra channel →