m2v-align

⚡ Fast bitext alignment powered by model2vec

Align parallel paragraphs/sentences between two languages in seconds — no GPU required.

Why m2v-align?

Traditional bitext aligners rely on large cross-lingual models (LaBSE, LASER) that are slow and memory-hungry. m2v-align uses Model2Vec static embeddings instead:

• 🚀 ~100x faster than transformer-based aligners — thousands of sentence pairs per second on CPU
• 🪶 Tiny footprint — models as small as 8M parameters, no PyTorch required at inference
• 🌍 Multilingual — works across 100+ languages with multilingual Model2Vec models
• 🔌 Simple API — align two lists of sentences in 3 lines of code

Installation

pip install m2v-align

Quickstart

Python API

from m2v_align import BitextAligner

aligner = BitextAligner()  # downloads a default multilingual model

source = [
    "Hello, how are you?",
    "The cat sits on the mat.",
    "I love programming.",
]
target = [
    "Ich liebe Programmieren.",
    "Hallo, wie geht es dir?",
    "Die Katze sitzt auf der Matte.",
]

pairs = aligner.align(source, target, threshold=0.6)

for pair in pairs:
    print(f"{pair.score:.3f} | {pair.source} <-> {pair.target}")

Output: