MrBERT is a multilingual foundational encoder model based on the ModernBERT architecture. It is pre-trained from scratch on a large-scale corpus of 6.1 trillion tokens covering 35 European languages as well as code. By building on ModernBERT’s modernized BERT-style design, MrBERT combines strong bidirectional representations with efficient long-context modeling.
Designed as a general-purpose multilingual encoder, MrBERT is well suited for a wide range of downstream tasks such as retrieval, classification, semantic search, and cross-lingual understanding across diverse languages.
Technical Description
Technical details of the MrBERT model.
Description
Value
Model Parameters
308M
Tokenizer Type
SPM
Vocabulary size
256,000
Precision
bfloat16
Context length
8192
Training Hyperparemeters
Hyperparameter
Value
Pretraining Objective
Masked Language Modeling
Learning Rate
1E-03
Learning Rate Scheduler
WSD
Warmup
3,000,000,000 tokens
Optimizer
decoupled_stableadamw
Optimizer Hyperparameters
AdamW (β1=0.9,β2=0.98,ε =1e-06 )
Weight Decay
1E-05
Global Batch Size
4096 (Short Context) / 512 (Long Context)
Dropout
1E-01
Activation Function
GeLU
How to use
You can use the pipeline for masked language modeling:
python
1>>>from transformers import pipeline
2>>>from pprint import pprint
3>>> unmasker = pipeline('fill-mask', model='BSC-LT/MrBERT')45>>> pprint(unmasker("I love the<mask>of Barcelona.",top_k=3))6[{'score':0.29333314299583435,7'sequence':'I love the city of Barcelona.',8'token':31489,9'token_str':'city'},10{'score':0.06682543456554413,11'sequence':'I love the capital of Barcelona.',12'token':10859,13'token_str':'capital'},14{'score':0.05594080686569214,15'sequence':'I love the streets of Barcelona.',16'token':178738,17'token_str':'streets'}]18>>> pprint(unmasker("Me encanta la<mask>de Barcelona.",top_k=3))19[{'score':0.4422685205936432,20'sequence':'Me encanta la ciudad de Barcelona.',21'token':19587,22'token_str':'ciudad'},23{'score':0.059732843190431595,24'sequence':'Me encanta la capital de Barcelona.',25'token':10859,26'token_str':'capital'},27{'score':0.03484857454895973,28'sequence':'Me encanta la arquitectura de Barcelona.',29'token':83374,30'token_str':'arquitectura'}]31>>> pprint(unmasker("M'encanta la<mask>de Barcelona.",top_k=3))32[{'score':0.45476993918418884,33'sequence':"M'encanta la ciutat de Barcelona.",34'token':17128,35'token_str':'ciutat'},36{'score':0.05597861483693123,37'sequence':"M'encanta la capital de Barcelona.",38'token':10859,39'token_str':'capital'},40{'score':0.04105329513549805,41'sequence':"M'encanta la música de Barcelona.",42'token':16051,43'token_str':'música'}]
Alternatively, you can also extract the logits associated with the sequences and perform the calculations by hand:
python
1from transformers import AutoTokenizer, AutoModelForMaskedLM
2import torch
34model = AutoModelForMaskedLM.from_pretrained("BSC-LT/MrBERT")5tokenizer = AutoTokenizer.from_pretrained("BSC-LT/MrBERT")67outputs = model(**tokenizer("The capital of Spain is<mask>", return_tensors="pt")).logits
89# The index of "<mask>" token is -2 given that the -1 position is the EOS token "</s>".10predicted_token = tokenizer.decode(torch.argmax(outputs[0,-2,:]))1112print(f"The decoded element is \"{predicted_token}\".")# This will give "Madrid"
Data
Pretraining Corpus
The pretraining corpus comprises 6.1 trillion tokens, covering 35 European languages and code. Training was conducted in three distinct phases to balance broad knowledge acquisition with long-context performance:
Phase
Context Length
Token Count
Short Context
1,024 tokens
5.5T
Long Context
8,192 tokens
500B
Annealing
8,192 tokens
100B
Language Distribution & Sources
While English constitutes 67.4% of the total data, the remaining distribution spans a diverse set of European languages as shown below:
drawing
The full list of datasets used will be published soon.
Multilingual Evaluation and Performance
Evaluation is done using multilingual benchmarks in order to assess the multilingual capabilities of the models.
The following multilingual benchmarks have been considered:
Benchmark
Description
Languages
Source
XTREME
Benchmark for the evaluation of the cross-lingual generalization ability of pre-trained multilingual models
NER Spanish tasks and Multilingual Text Embedding Benchmark for evaluating sentence and document embedding quality across retrieval tasks using ColBERT
Text Classification tasks and Multilingual Text Embedding Benchmark for evaluating sentence and document embedding quality across retrieval tasks using ColBERT
This section presents results across various multilingual benchmarks, with the maximum values highlighted in bold and the second-highest values underlined.
XTREME Benchmark
The Cross-lingual TRansfer Evaluation of Multilingual Encoders (XTREME) benchmark is designed to assess the cross-lingual generalization capabilities of pre-trained multilingual models. It comprises nine tasks that collectively test reasoning across various levels of syntax and semantics. The results reported here are from the test set, using the learning rate chosen based on the best-performing model on the validation set.
Given that retrieval tasks generally achieve higher performance with multi-vector methods such as ColBERT, we evaluate these tasks separately using m-TEB.
In the table below we just show the average results across languages:
task
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
xnli (TC)
78.25
79.09
80.54
77.90
81.26
pawsx (TC)
89.50
90.36
92.34
89.55
91.32
udpos (POS)
85.55
85.36
84.33
82.07
83.74
panx (NER)
73.69
75.65
73.89
73.05
72.06
tydiqa (QA)
56.41
53.96
63.95
51.07
56.34
mlqa (QA)
68.91
68.67
71.48
68.05
70.67
xquad (QA)
75.61
75.45
77.79
74.37
77.91
average
75.42
75.50
77.76
73.72
76.19
For a detailed description for each language, we provide the full table description:
🔵 Sentence Classification
🔵 XNLI
Metric used: Accuracy.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
bg
77.60
78.20
79.24
76.79
80.04
de
77.05
77.68
80.16
76.03
80.16
el
75.63
76.53
77.17
74.97
79.74
en
84.49
85.63
86.73
84.87
88.02
es
79.00
80.32
81.88
79.08
82.53
fr
78.04
78.94
80.76
78.68
80.78
ru
75.91
76.31
77.88
74.87
77.56
🔵 PAWS-X
Metric used: Accuracy.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
de
87.25
88.50
90.95
86.55
89.05
en
94.10
94.50
95.50
93.70
95.60
es
87.90
88.95
91.40
89.50
89.95
fr
88.75
89.50
91.50
88.45
90.70
🟣 Structured Prediction: POS
🟣 POS (UDPOS)
Metric used: F1.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
bg
88.31
88.10
87.48
86.33
86.90
de
88.35
87.54
86.94
85.52
86.82
el
87.52
83.90
81.80
83.39
83.95
en
95.89
95.80
95.96
95.65
95.84
es
87.49
87.45
87.70
86.03
87.84
et
84.60
85.55
81.08
79.52
77.43
eu
66.98
67.43
64.67
64.32
64.22
fi
84.77
83.58
82.38
79.01
77.05
fr
85.93
86.84
86.55
82.76
85.86
hu
83.12
82.15
80.00
77.72
81.69
it
86.95
89.02
89.14
85.73
87.42
lt
83.10
81.03
82.00
75.81
80.42
nl
89.16
89.36
89.17
86.66
89.02
pl
83.90
84.33
82.80
81.22
85.00
pt
86.75
87.78
87.35
85.07
87.28
ro
83.47
81.39
82.54
76.21
77.47
ru
88.92
89.59
87.84
85.48
88.75
uk
84.72
85.55
82.54
80.77
84.44
🟣 NER (PANX)
Metric used: F1.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
bg
76.42
77.92
74.29
73.79
74.97
de
74.60
78.36
75.98
75.59
76.11
el
73.32
76.31
68.62
68.95
70.88
en
82.39
83.43
82.82
83.99
82.05
es
71.04
81.19
79.93
78.14
77.46
et
71.85
72.87
70.32
69.92
67.04
eu
59.58
56.52
51.70
59.46
55.84
fi
75.80
76.07
74.67
74.78
70.54
fr
77.61
77.65
81.15
78.14
77.63
hu
76.66
73.42
73.03
74.14
73.26
it
76.73
80.02
80.49
79.65
76.98
lt
72.04
73.52
70.14
68.79
66.40
nl
79.85
81.91
80.10
78.87
79.37
pl
77.45
80.33
77.81
78.03
77.75
pt
75.93
79.21
80.36
78.86
77.36
ro
71.78
71.49
75.37
69.47
66.02
ru
63.39
67.31
64.13
61.73
60.89
uk
70.00
74.20
69.17
62.54
66.49
⚫ Question Answering
⚫ TyDiQA
Metric used: F1.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
en
62.42
60.39
71.73
57.94
68.50
fi
54.19
49.51
60.89
44.85
48.08
ru
52.61
51.98
59.24
50.42
52.45
⚫ MLQA
Metric used: F1.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
de
61.57
61.49
65.19
59.96
64.38
en
78.68
76.37
79.09
76.24
77.45
es
66.48
68.15
70.17
67.95
70.17
⚫ XQUAD
Metric used: F1.
langs
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
de
74.31
72.93
77.68
70.62
75.94
el
71.68
71.70
68.81
71.93
74.92
en
82.26
82.37
85.67
82.29
84.44
es
75.20
78.18
79.58
76.08
79.23
ru
74.58
72.08
77.20
70.93
75.02
EvalES Benchmark
The EvalES benchmark consists of 7 tasks: Named Entity Recognition and Classification (CoNLL-NERC), Part-of-Speech Tagging (UD-POS), Text Classification (MLDoc), Paraphrase Identification (PAWS-X), Semantic Textual Similarity (STS), Question Answering (SQAC), and Textual Entailment (XNLI). This benchmark evaluates the model's capabilities in the Spanish language.
tasks
xlm-roberta-base (279M)
mRoBERTa (283M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
MrBERT-es (150M)
pos (f1)
99.01
99.03
99.09
98.92
99.06
99.08
ner (f1)
86.91
87.77
87.01
86.96
87.42
87.77
sts (person)
80.88
79.69
82.88
84.52
84.18
85.23
tc - paws-x (acc)
90.35
91.30
91.35
89.70
91.25
91.90
tc - mldoc (acc)
47.67
91.28
95.10
96.13
95.28
95.55
tc - massivenew (acc)
21.89
86.45
86.79
87.19
87.46
87.05
qa (f1)
74.48
77.03
79.79
76.78
81.96
82.19
Average
71.60
87.51
88.86
88.60
89.52
89.83
CLUB Benchmark
The Catalan Language Understanding Benchmark consists of 6 tasks: Named Entity Recognition (NER), Part-of-Speech Tagging (POS), Semantic Textual Similarity (STS), Text Classification (TC), Textual Entailment (TE), and Question Answering (QA). This benchmark evaluates the model's capabilities in the Catalan language.
This comparison also includes RoBERTa-ca, a model derived from mRoBERTa by applying vocabulary adaptation and performing continual pre-training on a 95GB Catalan-only corpus. For further details, visit here.
tasks
xlm-roberta-base (279M)
mRoBERTa (283M)
roberta-ca (125M)
mmBERT (308M)
mGTE (306M)
MrBERT (308M)
MrBERT-ca (150M)
ner (F1)
87.61
88.33
89.70
88.14
87.20
87.32
88.04
pos (F1)
98.91
98.98
99.00
99.01
98.67
99.01
99.03
sts (Person)
74.67
79.52
82.99
83.16
78.65
83.00
85.42
tc (Acc.)
72.57
72.41
72.81
74.11
74.68
73.79
74.97
te (Acc.)
79.59
82.38
82.14
83.18
79.40
84.03
86.92
viquiquad (F1)
86.93
87.86
87.31
89.86
86.78
89.25
89.59
xquad (F1)
69.69
69.40
70.53
73.88
69.27
73.96
74.47
Average
81.42
82.70
83.50
84.48
82.09
84.34
85.49
MTEB Benchmark
Models are trained on 810k MS-Marco samples using teacher scores from BGE-M3, with a batch size of 16 and a learning rate of 8e-5, leveraging the PyLate library. Evaluations are conducted separately across the legal, scientific, and medical domains.
🩺 Biomedical
Task Name
Task Type
mmBERT (308M)
MrBERT (308M)
MrBERT-es (150M)
BioClinical-MdnBERT (150M)
Clinical MdnBERT (137M)
MrBERT-biomed (308M)
bsc-bio-distemist-ner (ES)
NER
78.00
77.84
78.07
75.45
70.22
77.93
cantemist (ES)
NER
78.03
68.73
73.40
66.68
30.91
70.78
pharmaconer (ES)
NER
89.66
88.58
88.97
87.66
81.69
89.92
AbSanitas (ES)
Retrieval
34.68
34.16
53.49
30.41
18.08
51.01
r2med (EN)
Retrieval
10.87
10.15
8.65
9.97
5.91
9.76
SciDocs (EN)
Retrieval
10.00
9.75
9.90
9.33
3.64
10.05
SciFact (EN)
Retrieval
32.35
31.08
31.46
32.07
20.34
30.25
TREC-COVID (EN)
Retrieval
30.77
49.53
37.51
46.08
23.88
48.76
Average (EN)
All Tasks
21.00
25.13
21.88
24.36
13.44
24.71
Average (EN + ES)
All Tasks
45.55
46.23
47.68
44.71
31.83
48.56
⚖️ Legal
Task Name
Task Type
mmBERT (308M)
MrBERT (308M)
MrBERT-es (150M)
legal-bert- base-uncased (110M)
MrBERT-legal (308M)
LexBOE (ES)
Text Classification
96.84
97.02
97.28
95.36
96.80
small-spanish-legal-dataset (ES)
Retrieval
42.58
40.78
46.92
19.79
38.75
EURLEX (EN)
Text Classification
97.43
97.40
97.41
97.42
97.33
AILAStatutes (EN)
Retrieval
14.31
13.90
12.28
13.49
16.33
legal_summarization (EN)
Retrieval
53.33
53.84
46.41
52.40
55.05
LegalBench (EN)
Retrieval
60.15
58.88
58.26
63.42
58.04
NanoTouche2020 (EN)
Retrieval
34.03
44.15
31.18
34.48
44.74
Average (EN)
All Tasks
51.85
53.63
49.11
52.24
54.30
Average (EN + ES)
All Tasks
56.95
58.00
55.68
53.77
58.15
Additional information
Author
The Language Technologies Lab from Barcelona Supercomputing Center.
Contact
For further information, please send an email to langtech@bsc.es.
Copyright
Copyright(c) 2026 by Language Technologies Lab, Barcelona Supercomputing Center.
Funding
This work has been supported and funded by the Ministerio para la Transformación Digital y de la Función Pública and the Plan de Recuperación, Transformación y Resiliencia – funded by the EU through NextGenerationEU, within the framework of the Modelos del Lenguaje project, as well as by the European Union – NextGenerationEU. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Commission. Neither the European Union nor the European Commission can be held responsible for them.
Acknowledgements
This project has benefited from the contributions of numerous teams and institutions through data contributions.
In Catalonia, many institutions have been involved in the project. Our thanks to Òmnium Cultural, Parlament de Catalunya, Institut d'Estudis Aranesos, Racó Català, Vilaweb, ACN, Nació Digital, El món and Aquí Berguedà.
At national level, we are especially grateful to our ILENIA project partners: CENID, HiTZ and CiTIUS for their participation. We also extend our genuine gratitude to the Spanish Senate and Congress, Fundación Dialnet, Fundación Elcano, the "Instituto de Ingenieria del Conocimiento" and the ‘Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en Ingeniería (SIANI)’ of the University of Las Palmas de Gran Canaria.
At the international level, we thank the Welsh government, DFKI, Occiglot project, especially Malte Ostendorff, and The Common Crawl Foundation, especially Pedro Ortiz, for their collaboration.
Their valuable efforts have been instrumental in the development of this work.
Disclaimer
Be aware that the model may contain biases or other unintended distortions.
When third parties deploy systems or provide services based on this model, or use the model themselves,
they bear the responsibility for mitigating any associated risks and ensuring compliance with applicable regulations,
including those governing the use of Artificial Intelligence.
The Barcelona Supercomputing Center, as the owner and creator of the model, shall not be held liable for any outcomes resulting from third-party use.
Citation
@article{tamayo2026mrbert,
title={MrBERT: Modern Multilingual Encoders via Vocabulary, Domain, and Dimensional Adaptation},
author={Tamayo, Daniel and Lacunza, I{\~n}aki and Rivera-Hidalgo, Paula and Da Dalt, Severino and Aula-Blasco, Javier and Gonzalez-Agirre, Aitor and Villegas, Marta},
journal={arXiv preprint arXiv:2602.21379},
year={2026}
}
