NLLB-200 3.3B

NLLB-200 3.3B

Translates single sentences among 200 languages, with a focus on low-resource ones. A research model from facebook for machine translation researchers, not meant for production or certified translations.

Made by facebook 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: NLLB-200 is a machine translation model that translates single sentences among 200 languages. This card covers the 3.3B variant. It was built mainly for research, with a focus on languages that have little training text.

Who it suits: It suits machine translation researchers and teams who study low-resource languages. The card names researchers as its main users.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Research | Machine translation researchers | Study sentence translation across many languages | Compare a language pair against a baseline
Language technology | Teams working on low-resource languages | Try single-sentence translation for rarely supported languages | Prototype a translator for one language
Software | NLP engineers | Test the model in the transformers library before a real build | A quick local trial on short sentences

If you need document translation, medical or legal text, or certified translations, you can skip it.

Where to use it:

• Windows, macOS or Linux with Python: for downloading the files
• The transformers library, the library tag on the model page
• The Fairseq code repository, where the card says usage and training code live

Title: Jobs it does
Use case | Who | What they get back
Single-sentence translation | Researchers | One translated sentence in the target language
Low-resource language study | Language technology teams | Output to inspect for languages with little data

How to set it up:

• Python on Windows, macOS or Linux
• Enough disk space for a 3.3B model

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Follow the Fairseq repository for usage, since the card sends you there.

pip install -U huggingface_hub
hf download facebook/nllb-200-3.3B

How to use it: Translate one sentence at a time between two of its 200 languages. The card says the model was trained on inputs up to 512 tokens, so longer text may translate worse. It sends you to the Fairseq repository for the how-to.

hf download facebook/nllb-200-3.3B

Example: You ask it to translate a short sentence, such as "Where is the station?", into one of its 200 languages. A single translated sentence comes back. The card warns that mistranslations can remain, so check anything important.

For newcomers and experts: If you are new, machine translation means a program turns text from one language into another. The first small thing to try is the download command, then one short sentence in the Fairseq instructions.

If you know the field, the card says the model learned from parallel multilingual data plus monolingual text built from Common Crawl. It was scored with BLEU, spBLEU and chrF++, plus human XSTS ratings and a toxicity check, on Flores-200. The trade-off is range: tested mostly on Wikimedia text, with limited checks on other domains.

A sentence in the source language
SentencePiece splits it into pieces
NLLB-200 3.3B translates it
A sentence in the target language

Good to know:

• Research use: it is meant for experiments and studies, and it should not be put into a live production service.
• Domain: it was trained on general text, not on specialist texts such as medical or legal.
• Length: it was trained on inputs up to 512 tokens, and document translation is out of scope.
• Certification: its output cannot stand in for a certified translation.
• Risk: mistranslations may remain, which matters when people rely on them for health or safety decisions.
• Training data: it came from public web sources, and personal information may not be fully removed.
• License: cc-by-nc-4.0, so it is non-commercial.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: you download files from Hugging Face, and the card names no other access.
• Not for: production deployment, document translation, specialist domains and certified translations.

On Hugging Face, 13 November 2026

• Task: translation
• Library: transformers
• Downloads in the last 30 days: 293,185
• Likes: 493
• License: cc-by-nc-4.0
• Languages: ace, acm, acq, aeb, af, ajp, ak, als
• Quantized versions: 9
• Fine-tunes: 39
• Tags: transformers, pytorch, m2m_100, text2text-generation, nllb, translation

Links

• Model: https://huggingface.co/facebook/nllb-200-3.3B
• Files: https://huggingface.co/facebook/nllb-200-3.3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download facebook/nllb-200-3.3B

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