SMaLL-100

SMaLL-100

Translates text between a long list of languages, covering more than 10K language pairs, with a model much smaller and faster than M2M-100. Made for developers and researchers who need compact multilingual translation.

Made by alirezamsh 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: SMaLL-100 is a compact machine translation model that covers more than 10K language pairs. The card says it matches M2M-100 on competitive results while being much smaller and faster. It was introduced in an EMNLP 2022 paper and uses the same architecture and config as M2M-100.

Who it suits: It suits developers, localization teams and researchers who want a small translation model that handles many languages, including low-resource ones.

Industry | Who uses it | How they use it | Example
Software | App developers | Add on-device or self-hosted translation to a product | Translate a user message from Hindi to French
Localization | Translation teams | Draft translations across many language pairs | Convert Chinese text into English
Research | Machine translation researchers | Study shallow models for low-resource languages | Compare scores against M2M-100
Education | Language learners and teachers | See quick translations between languages | Check a sentence in another language

If you need translation of whole documents with layout, or a hosted service, you can skip it.

Where to use it: It runs in Python through the transformers library on your own computer.

• Python scripts using transformers
• Windows, macOS or Linux
• Research setups that compute spBLEU with the original repository

Use case | Who | What they get back
Translate a sentence | Developers | Text in the target language
Supervised training | Researchers | A loss value from a source and target pair
Low-resource translation | Language teams | Translations for languages large models cover poorly

How to set it up: You need Python with transformers and the sentencepiece package. The tokenizer is modified, so the card says to load it locally from the tokenization_small100.py file in the repository.

1. Install the Hugging Face hub tools.
2. Download the model files, which include the tokenizer file.
3. Install sentencepiece.
4. Import the SMALL100Tokenizer from that local file.

pip install -U huggingface_hub
hf download alirezamsh/small100

pip install sentencepiece

How to use it: You set the target language code on the tokenizer, encode your text, generate, and decode. The card uses a beam size of 5 and a maximum target length of 256 for generation.

from tokenization_small100 import SMALL100Tokenizer

hi_text = "जीवन एक चॉकलेट बॉक्स की तरह है।"
chinese_text = "生活就像一盒巧克力。"

tokenizer = SMALL100Tokenizer.from_pretrained("alirezamsh/small100")

# translate Hindi to French
tokenizer.tgt_lang = "fr"
encoded_hi = tokenizer(hi_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)

Put your own value in its key setting first.

Tool | What it does
tokenizer.tgt_lang | Sets the target language code, such as fr or en
model.generate | Produces the translated tokens
tokenizer.batch_decode | Turns the tokens back into text
hf download | Fetches the model files to your computer

Example: You take the Chinese sentence for life is like a box of chocolates, set the target language to en, and run generate. The card shows the English result as "Life is like a box of chocolate." A Hindi version of the same sentence goes to French the same way.

For newcomers and experts: A translation model reads text in one language and writes it in another. To try it, run the Hindi to French example above and then change the language code.

The network is the M2M-100 design, and only the tokenizer is changed to handle language codes. The input is the target language code, then the source tokens and an end token. Supervised training feeds source and target text and returns a loss.

Set the target language code
Tokenize the source text
Generate with beam size 5
Decode tokens to translated text

Good to know: The language list runs from Afrikaans and Amharic to Yoruba and Zulu, with codes such as fr, en, zh and hi.

• Tokenizer: it must be loaded from the local tokenization_small100.py file for now.
• Evaluation: spBLEU is computed with the original repository.
• Training data: the card says it can be provided on request.
• License: MIT.
• Cost to run: none named.
• Data it reaches: the text you translate, handled on your own machine.
• Not for: the card names no exclusions.

On Hugging Face, 13 November 2026

• Task: translation
• Library: transformers
• Parameters: 333M
• Weights: about 1.3 GB in F32; about 166 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 10,902
• Likes: 113
• License: mit
• Languages: af, am, ar, ast, az, ba, be, bg
• Quantized versions: 4
• Fine-tunes: 4
• Tags: transformers, pytorch, onnx, safetensors, m2m_100, text2text-generation, small100, translation, flores101, tico19, tatoeba, multilingual

Links

• Model: https://huggingface.co/alirezamsh/small100
• Files: https://huggingface.co/alirezamsh/small100/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download alirezamsh/small100

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