Train 400x faster Static Embedding Models with Sentence Transformers

Hugging Face just took embeddings back to their roots. The company released two new models — one for English retrieval, one for multilingual similarity — that ditch the attention layers powering nearly every modern embedding model and replace them with something almost embarrassingly simple: a lookup table. Type in a word, get back a pre-computed vector. No context window, no self-attention, no waiting.

That sounds like a step backward, and in a sense it is. Static embeddings like GloVe and word2vec predate transformers by years. What's changed is how Hugging Face trained them. Using contrastive learning — the same technique that pulls similar sentence pairs together and pushes dissimilar ones apart — the team squeezed real gains out of an architecture everyone assumed was obsolete. The retrieval model, static-retrieval-mrl-en-v1, and the similarity model, static-similarity-mrl-multilingual-v1, run 100x to 400x faster on CPU than comparable transformer models like all-mpnet-base-v2 and multilingual-e5-small. And they still hold onto at least 85% of the quality those heavier models deliver.

The training recipe leaned on 30 datasets for training and 13 for evaluation, deliberately avoiding overlap with the BEIR retrieval benchmark so the results would reflect genuine zero-shot performance rather than memorization. For the English model, the team used the bert-base-uncased tokenizer; for the multilingual one, bert-base-multilingual-uncased. Both produce 1024-dimensional vectors, matching the output size of larger models like bge-large-en-v1.5. Layered on top is Matryoshka Representation Learning, a trick that trains the model to pack the most useful information into the first few dimensions of the embedding, so developers can truncate vectors for even faster downstream tasks without much quality loss.

What makes this interesting isn't just the speed number — it's where that speed becomes useful. Running semantic search entirely in a browser tab, on a Raspberry Pi, or inside a mobile app without touching a GPU is suddenly realistic rather than aspirational. Hugging Face built the models to drop straight into the existing Sentence Transformers workflow, so switching from a slow transformer-based encoder to one of these takes a one-line change, not a rewrite.

The team is upfront that this is a starting point, not a finished product. They've flagged data quality improvements and further architectural tweaks as open problems for the community to chase. But the core claim — that you can get transformer-adjacent quality out of a glorified dictionary lookup — is the kind of result that quietly reshapes what people think is required to ship an embedding model.