statembed

Fast, lightweight static text embeddings.

statembed loads pre-trained static embedding models stored in the Safetensors format and produces fixed-length text embeddings via mean-pooling over token-level vectors, with optional L2 normalization. There's no neural network forward pass at inference time (just a lookup-and-average over an embedding matrix) which makes it extremely fast compared to transformer-based embedding models, at some cost to embedding quality.

This repository is a Cargo workspace with two crates:

How it works

1. Lazy loading — the model tensor (and tokenizer, in Rust) are loaded from disk on the first embedding call, then cached in memory.
2. Mean pooling — each token ID maps to a row in the embedding matrix; the rows for all tokens in the input are averaged into a single fixed-length vector.
3. Optional normalization — when enabled, the resulting vector is L2-normalized.
4. Token caching — decoded token vectors are cached so repeated tokens across calls don't need to be re-decoded.

Rust: statembed

[dependencies]
statembed = "0.1"

use statembed::StaticEmbedding;

let mut model = StaticEmbedding::from_dir("./my-model", Some(true))?;
let embedding = model.embed_text("hello world", None)?;

The model directory must contain model.safetensors, plus tokenizer.json if using the (default) tokenizers feature to embed raw text. Models can also be downloaded directly from the Hugging Face Hub with the hf-hub feature. See crates/statembed/README.md for the full feature list (tokenizers, hf-hub, mmap, simd) and API details.

Python: statembed-py

uv add statembed-py

Python bindings embed pre-tokenized input only, so you need to pair them with a tokenizer library such as tokenizers:

from tokenizers import Tokenizer
from statembed_py import StaticEmbedding

model = StaticEmbedding(model_dir="./my-model")  # must contain model.safetensors
tokenizer = Tokenizer.from_file("./my-model/tokenizer.json")

tokens = tokenizer.encode("hello world").ids
embedding = model.embed_tokens(tokens)

See crates/statembed-py/README.md for installation options, building from source, and the full API.

Development

This is a Cargo workspace (resolver = "3"); run standard Cargo commands from the repo root:

cargo build
cargo test
cargo bench -p statembed   # compares against model2vec-rs

License

MIT