BERT multilingual base model (cased)

BERT multilingual base model (cased)

Fills in masked words and extracts text features in 104 languages. A case-sensitive BERT base model meant to be fine-tuned for classification, tagging and question answering.

Made by google-bert on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a multilingual BERT base model that Google pretrained on the Wikipedias of 104 languages. It learned by guessing masked words and by judging whether two sentences follow each other. It is case sensitive, so english and English count as different words.

Who it suits: It suits developers and researchers who fine-tune a model for a task that reads whole sentences, in many languages. People have published 1,038 fine-tuned versions of it.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | NLP developers | Fine-tune it for sequence or token classification | A topic tagger for text in several languages
Research | Language researchers | Use it as a multilingual baseline | Compare fine-tuned versions on one task
Support | Teams handling multilingual tickets | Build a classifier on top of its features | Sort tickets by topic across languages

If you want a model that writes text, you can skip it. The card points to GPT2-style models for that.

Where to use it:

• Python with the transformers library, on Windows, macOS or Linux
• PyTorch or TensorFlow, since the card shows both
• Fine-tuning projects for classification, token classification and question answering

Title: Jobs it does
Use case | Who | What they get back
Masked word prediction | Developers | Ranked guesses for the hidden word
Text features | Researchers | Feature vectors you can feed to a classifier
Task fine-tuning | NLP teams | A base to train for classification or question answering

How to set it up:

• Python with the transformers library
• PyTorch for the feature example, or TensorFlow for the other

1. Install transformers and PyTorch.
2. Optionally download the model files by hand.
3. Load the model by name in your code.

pip install -U huggingface_hub
hf download google-bert/bert-base-multilingual-cased

How to use it: Use the fill-mask pipeline for a quick test, or load the model to get features for a text. The card says the raw model mainly serves as a starting point for fine-tuning.

>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-cased')
>>> unmasker("Hello I'm a [MASK] model.")

from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased')
model = BertModel.from_pretrained("bert-base-multilingual-cased")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

Example: You give the pipeline the sentence "Hello I'm a [MASK] model." and it returns five guesses for the hidden word. In the card's output the top guesses are model, world and data, with scores near 0.10, 0.05 and 0.05.

For newcomers and experts: If you are new, think of it as a model that learned languages by filling in blanks. The first small thing to try is the fill-mask pipeline with a sentence that has one hidden word.

If you know the field, the model uses a masked objective that hides 15% of the words, plus next sentence prediction, and a WordPiece vocabulary of 110,000 shared across languages. The trade-off is fit: it is built for tasks on whole sentences, not for generating text.

Hide 15% of the words in a text
Run the whole sentence through BERT
Predict the hidden words
Reuse the features for a task

Good to know:

• Language: pretraining covered 104 languages, picked as those with the biggest Wikipedia editions.
• Intended use: fine-tuning on whole-sentence tasks such as classification and question answering.
• Case: the model is case sensitive.
• Size: 179M parameters, about 714 MB in F32.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: you download the model from Hugging Face, and the card names no other access.
• Not for: text generation, where the card suggests a model like GPT2.

On Hugging Face, 13 November 2026

• Task: fill mask
• Library: transformers
• Parameters: 179M
• Weights: about 714 MB in F32; about 89 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,539,340
• Likes: 623
• License: apache-2.0
• Languages: af, sq, ar, an, hy, ast, az, ba
• Quantized versions: 11
• Fine-tunes: 1,043
• Tags: transformers, pytorch, safetensors, bert, fill-mask, multilingual

Links

• Model: https://huggingface.co/google-bert/bert-base-multilingual-cased
• Files: https://huggingface.co/google-bert/bert-base-multilingual-cased/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download google-bert/bert-base-multilingual-cased

Then follow the usage in its model card. Nothing in this kit runs by itself.