Try LFM • Documentation • LEAP

LFM2.5-Embedding-350M

LFM2.5-Embedding-350M is a dense bi-encoder for fast multilingual retrieval. It produces a single vector per document — the smallest, fastest index — for reliable cross-lingual search across 11 languages.

Best-in-class multilingual accuracy for a dense embedder of its size.
Inference speed is on par with much smaller models, thanks to the efficient LFM2 backbone.
You can use it as a drop-in replacement in your current RAG pipelines.

Find more information about LFM2.5-Embedding-350M in our blog post.

🏃 How to run

Example usage with llama.cpp:

Start llama-server

llama-server -hf LiquidAI/LFM2.5-Embedding-350M-GGUF --embeddings

Make requests to embed queries and documents, and rank by cosine similarity (note the asymmetric query: / document: prompt prefixes)

❯ uv run dense-retrieve.py

Score: -0.1783 | Q: What is panda? | D: hi
Score:  0.0511 | Q: What is panda? | D: it is a bear
Score:  0.5657 | Q: What is panda? | D: The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.

# /// script
# requires-python = ">=3.10"
# dependencies = ["numpy", "requests"]
# ///

# dense-retrieve.py
import numpy as np, requests

QUERY_PREFIX, DOC_PREFIX = "query: ", "document: "

def embed(text: str) -> np.ndarray:
    r = requests.post(
        "http://localhost:8080/v1/embeddings",
        json={"input": text},
    )
    v = np.array(r.json()["data"][0]["embedding"])
    return v / np.linalg.norm(v)

docs = [
    "hi",
    "it is a bear",
    "The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China.",
]
query = "What is panda?"

q = embed(QUERY_PREFIX + query)
for doc in docs:
    d = embed(DOC_PREFIX + doc)
    print(f"Score: {float(q @ d):.4f} | Q: {query} | D: {doc}")