Index-Translate-2B

Index-Translate-2B

Translates text between 150 languages and follows instructions about glossaries, formatting, tone and length. A compact 2.3B model from IndexTeam for teams that translate locally, with image translation through companion files.

Made by IndexTeam 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: Index-Translate-2B is a small multilingual translation model from IndexTeam, built on Qwen3.5. It translates between 150 languages and obeys written instructions about terms, layout, tone and output length. It can also read text inside images and translate it without a separate OCR tool, using companion files called mmproj.

Who it suits: Translators, localization teams and developers who want a compact model they can run themselves, and who need control over glossaries, tone and file formats. The files carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Localization | Translation teams | Translate product text with a fixed glossary | Keeping approved brand terms the same in every language
Media | Subtitle and article teams | Translate subtitles and articles into many languages | Turning an English article into several target languages
Software | Developers | Translate strings that hold placeholders or structured data | Translating a JSON file without breaking its keys
Social platforms | Community managers | Translate slang and memes with their intended meaning | Translating a playful post so the joke survives

Anyone who needs speech or long-document translation can skip it, since those come as separate models in the same family.

Where to use it:

• Windows, macOS or Linux with Python, for downloading the files
• A local setup that loads the weights, which are about 4.5 GB in BF16
• The team's online demo, for a quick try before you download anything

Use case | Who | What they get back
Glossary-bound translation | Localization teams | Text in the target language that keeps the terms you require
Structured text translation | Developers | JSON, CSV or markdown with code and placeholders left alone
Style-aware translation | Content teams | A formal, casual or social-media tone in the output
Image translation | App builders | Text read from an image and translated in one step through the mmproj files

How to set it up:

• Python and the huggingface_hub package
• Room for the weights, about 4.5 GB in BF16, or about 1.1 GB at 4-bit as an estimate
• For image translation, the mmproj companion files

1. Install Python if you do not have it yet.
2. Install the huggingface_hub package with the first command below.
3. Download the model files with the second command.
4. Follow the model card for the loading steps and the mmproj files.

pip install -U huggingface_hub
hf download IndexTeam/Index-Translate-2B

How to use it: Once the files are on your computer, you load the model and send it the text plus a plain instruction. The instruction can name the target language, a glossary, a tone or a length limit. The card text available here shows no run command, so use the model page for the loading steps.

Tool | What it does
pip install -U huggingface_hub | Installs the download command
hf download IndexTeam/Index-Translate-2B | Fetches the model files to your computer
mmproj | Lets the model translate text found inside images

Example: You ask for a Japanese translation of an app message that holds a name placeholder and an order number in brackets, and you tell the model to leave both unchanged. The card says placeholders like these are kept.

For newcomers and experts: A translation model turns text from one language into another. A good first step is to download the files and read the model page. Then try one short sentence with a clear instruction.

The model has 2.3B parameters on Hugging Face, though the card calls it a 2B model. Training used a shared multilingual stage of 167.77B tokens, then three specialists for general translation, instruction following and memes. They are merged with weights 0.8, 0.1 and 0.1. A final distillation step with several teacher models patches task types that stay weak.

Download the files with hf
Load the model and any mmproj files
Send text plus an instruction
Read the translation

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the download fetches files from Hugging Face, and the source names nothing else
• The scores below come from the card's technical report, on its own data and settings
• On low-resource instruction translation, the card reports 3.97 percent off-target outputs
• Hugging Face lists 23 quantized versions and 3 fine-tuned versions made by other people

Title: Reported scores, higher is better
instTrans IFscore: 0.7569 score
MEME: 0.6443 score
Low-resource IFscore: 0.6586 score
Low-resource quality: 0.3050 score

Title: Weights size in GB
BF16: 4.5 GB
4-bit estimate: 1.1 GB

On Hugging Face, 13 November 2026

• Task: translation
• Parameters: 2.3B
• Weights: about 4.5 GB in BF16; about 1.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,891
• Likes: 81
• License: apache-2.0
• Quantized versions: 24
• Fine-tunes: 3
• Tags: safetensors, qwen3_5, translation, index

Links

• Model: https://huggingface.co/IndexTeam/Index-Translate-2B
• Files: https://huggingface.co/IndexTeam/Index-Translate-2B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download IndexTeam/Index-Translate-2B

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