This Model2Vec model is a distilled version of ibm-granite/granite-embedding-97m-multilingual-r2 trained for multilingual retrieval tasks. It uses static embeddings, allowing text embeddings to be computed orders of magnitude faster on both GPU and CPU. It is designed for applications where computational resources are limited or where real-time performance is critical.

Preliminary Metrics

Numbers for base model and popular static models are presented for context.

Model Info

model	vocab x dims	params	Size on Disk
ibm-granite/granite-embedding-97m-multilingual-r2	179,936 x 384	97M	211M
potion-retrieval-32M	63,091 x 512	32.3M	125M
potion-multilingual-128M	500,353 x 256	128.1M	1003M
static-similarity-mrl-multilingual-v1	105,879 x 1024	108.4M	417M
static-retrieval-multilingual-69m-v1	179,936 x 384	69.1M	274M

Embedding Performance

This is measured on a small test: 5000 texts, 354 chars/doc, 34 chars/query, 4 CPU cores, best of 10 runs.

model	docs/s	queries/s
granite-embedding-97m-multilingual-r2	7	69
potion-retrieval-32M	8431	39632
potion-multilingual-128M	5148	37727
static-similarity-mrl-multilingual-v1	7924	37731
static-retrieval-multilingual-69m-v1	8974	44456

NanoBEIR Multilingual Results

The scores below are averages per language by model. The best value among static models is highlighted.

NOTE: When looking at the scores below it's important to keep in mind that potion-retrieval-32M is an English model and two multilingual static models (potion-multilingual-128M and static-similarity-mrl-multilingual-v1) were not trained for retrieval.

Language	granite-embedding-97m-multilingual-r2	potion-retrieval-32M	potion-multilingual-128M	static-similarity-mrl-multilingual-v1	static-retrieval-multilingual-69m-v1
ara-Arab	0.4589	0.0988	0.2692	0.2860	0.3458
deu-Latn	0.5338	0.2537	0.3273	0.3454	0.4054
eng-Latn	0.5881	0.5107	0.3696	0.4352	0.4700
fra-Latn	0.5318	0.2828	0.3472	0.3702	0.4102
ita-Latn	0.5176	0.2752	0.3401	0.3683	0.3837
jpn-Jpan	0.4956	0.1142	0.3016	0.3190	0.3597
kor-Kore	0.4927	0.1171	0.3046	0.2763	0.3023
nor-Latn	0.4827	0.2532	0.3190	0.3314	0.3059
por-Latn	0.5193	0.2682	0.3364	0.3743	0.3992
spa-Latn	0.5295	0.2542	0.3369	0.3754	0.4145
swe-Latn	0.4991	0.2674	0.3154	0.3408	0.3208
Average	0.5136	0.2450	0.3243	0.3475	0.3743

RTEB

The best value in each test is highlighted.

Test	Language	potion-retrieval-32M	potion-multilingual-128M	static-similarity-mrl-multilingual-v1	static-retrieval-multilingual-69m-v1
AILACasedocs	eng-Latn	0.2157	0.2037	0.2202	0.2231
AILAStatutes	eng-Latn	0.1901	0.1598	0.1663	0.2100
AppsRetrieval	eng-Latn, python-Code	0.0431	0.0366	0.0127	0.0321
ChatDoctorRetrieval	eng-Latn	0.2470	0.1362	0.1535	0.2443
CUREv1	eng-Latn, eng-Latn	0.3019	0.2152	0.2548	0.2984
	fra-Latn, eng-Latn	0.0639	0.1325	0.1716	0.1694
	spa-Latn, eng-Latn	0.0276	0.1359	0.1615	0.1678
DS1000Retrieval	eng-Latn, python-Code	0.2330	0.2037	0.2083	0.1595
FinanceBenchRetrieval	eng-Latn	0.3571	0.2626	0.2829	0.2580
FinQARetrieval	eng-Latn	0.4905	0.4395	0.4096	0.4067
FreshStackRetrieval	eng-Latn, python-Code, javascript-Code, go-Code	0.2005	0.1725	0.1654	0.1773
HC3FinanceRetrieval	eng-Latn	0.2701	0.1952	0.2025	0.3653
HumanEvalRetrieval	eng-Latn, python-Code	0.4271	0.3738	0.3461	0.3398
LegalQuAD	deu-Latn	0.3917	0.4326	0.4110	0.3707
LegalSummarization	eng-Latn	0.5473	0.5286	0.5496	0.5378
MBPPRetrieval	eng-Latn, python-Code	0.2585	0.2464	0.2624	0.2156
MIRACLRetrievalHardNegatives	ara-Arab	0.0413	0.1657	0.1971	0.3515
	ben-Beng	0.0168	0.2388	0.2118	0.4738
	deu-Latn	0.1051	0.1268	0.1594	0.2459
	eng-Latn	0.2658	0.1391	0.1874	0.2446
	fas-Arab	0.0259	0.1464	0.1642	0.2909
	fin-Latn	0.2243	0.1501	0.2627	0.4206
	fra-Latn	0.0936	0.2027	0.1492	0.2307
	hin-Deva	0.0281	0.1773	0.1710	0.3357
	ind-Latn	0.1317	0.2174	0.1871	0.3355
	jpn-Jpan	0.0501	0.1261	0.1790	0.2826
	kor-Kore	0.0898	0.1178	0.2405	0.3905
	rus-Cyrl	0.0246	0.2299	0.1681	0.2592
	spa-Latn	0.1317	0.1838	0.2110	0.2701
	swa-Latn	0.1740	0.1478	0.2319	0.5065
	tel-Telu	0.0004	0.2628	0.1098	0.4842
	tha-Thai	0.0083	0.2577	0.0149	0.4425
	yor-Latn	0.2811	0.2853	0.3673	0.3509
	zho-Hans	0.0290	0.1742	0.1790	0.2497
SWEbenchCodeRetrieval	eng-Latn, python-Code	0.0288	0.0243	0.0255	0.0227
WikiSQLRetrieval	eng-Latn, sql-Code	0.3465	0.1759	0.1512	0.1834
Average		0.1560	0.1858	0.2034	0.2561

Training

Stage	Details
Base model	ibm-granite/granite-embedding-97m-multilingual-r2
Pre-Training	C4. 101 languages: 'af', 'am', 'ar', 'az', 'be', 'bg', 'bg-Latn', 'bn', 'ca', 'ceb', 'co', 'cs', 'cy', 'da', 'de', 'el', 'el-Latn', 'en', 'eo', 'es', 'et', 'eu', 'fa', 'fi', 'fil', 'fr', 'fy', 'ga', 'gd', 'gl', 'gu', 'ha', 'haw', 'hi', 'hi-Latn', 'hmn', 'ht', 'hu', 'hy', 'id', 'ig', 'is', 'it', 'iw', 'ja', 'ja-Latn', 'jv', 'ka', 'kk', 'km', 'kn', 'ko', 'ku', 'ky', 'la', 'lb', 'lo', 'lt', 'lv', 'mg', 'mi', 'mk', 'ml', 'mn', 'mr', 'ms', 'mt', 'my', 'ne', 'nl', 'no', 'ny', 'pa', 'pl', 'ps', 'pt', 'ro', 'ru', 'ru-Latn', 'sd', 'si', 'sk', 'sl', 'sm', 'sn', 'so', 'sq', 'sr', 'st', 'su', 'sv', 'sw', 'ta', 'te', 'tg', 'th', 'tr', 'uk', 'ur', 'uz', 'vi', 'xh', 'yi', 'yo', 'zh', 'zh-Latn', 'zu'
Fine-Tuning	mMARCO ("arabic", "chinese", "dutch", "english", "french", "german", "hindi", "indonesian", "italian", "japanese", "portuguese", "russian", "spanish", "vietnamese"), GooAQ, S2ORC, Free-Law-Project/opinions-synthetic-query-512, FIQA, MIRACL ('ar', 'bn', 'en', 'es', 'fa', 'fi', 'fr', 'hi', 'id', 'ja', 'ko', 'ru', 'sw', 'te', 'th', 'zh')

Installation

Install model2vec using pip:

pip install model2vec

Usage

Using Model2Vec

The Model2Vec library is the fastest and most lightweight way to run Model2Vec models.

Load this model using the from_pretrained method:

python

from model2vec import StaticModel

# Load a pretrained Model2Vec model
model = StaticModel.from_pretrained("amgix/static-retrieval-multilingual-69m-v1")

# Compute text embeddings
embeddings = model.encode(["Example sentence"])

Using Sentence Transformers

You can also use the Sentence Transformers library to load and use the model: