SalamandraTA-7b-instruct

SalamandraTA-7b-instruct

A 7.8B translation model for 40 languages and 3 varieties, covering sentence, paragraph and document translation plus HTML or XML text. For developers and researchers who want a local translator.

Made by BSC-LT 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: SalamandraTA-7b-instruct is a translation model from BSC-LT, tuned with instructions from a base model that stays internal. It translates in 40 languages plus 3 varieties, including Arabic, Japanese, Hindi, Korean and Simplified Chinese. Version 3 also handles terminology-aware translation, structured text, post-editing and named entity recognition.

Who it suits: It suits developers, researchers and companies who need machine translation they can run on their own hardware. The card allows research as well as commercial use of the model.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Software | Application developers | Add translation to a product or pipeline | Translate a paragraph from Spanish to Japanese
Web publishing | Content teams | Translate pages that contain HTML or XML | Keep the markup while translating the text
Language services | Translators and editors | Post-edit machine output and keep terms consistent | Fix a draft translation against a glossary

Anyone who wants a chat assistant can skip it, since the card says it has no chat abilities.

Where to use it:

• Python with the transformers library
• Windows, macOS or Linux, to download the files
• A GPU or large-memory machine for a 7.8B model

Title: Top jobs
Use case | Who | What they get back
General translation | Developers | Translated sentences, paragraphs or whole documents
Structured text | Content teams | Translated HTML or XML text
Post-editing | Editors | An improved version of a draft translation
Term-aware translation | Language teams | A translation that follows a given terminology

How to set it up:

• Python on your computer
• The huggingface_hub package
• About 15.5 GB of disk for the BF16 weights, or about 3.9 GB at 4-bit by estimate

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load the model with the transformers library, as in the card's example.

pip install -U huggingface_hub
hf download BSC-LT/salamandraTA-7b-instruct

How to use it: Load the model with AutoModelForCausalLM from transformers. The current release is on the main branch. To pick an older release, pass the revision argument when loading.

model = AutoModelForCausalLM.from_pretrained(
    "BSC-LT/salamandraTA-7b-instruct",
    revision="v1.0"  # omit revision for current release
)

Title: Releases
Version | Where to find it | Status
v3.0 | main branch | Current release
v2.0 | v2.0 revision | Second release
v1.0 | v1.0 revision | Initial release

Example: You load the current release and ask it to translate a paragraph from Catalan into English. The model returns the translation as text. For other tasks, such as post-editing, the card's usage section has the prompt format to follow.

For newcomers and experts: If you are new, think of this as a translation engine rather than a chatbot. For a first try, translate one short paragraph and read the result against the original.

For experts, it is a 32-layer decoder with 7,768,117,248 parameters and a context length of 8,192 tokens. The base was continually pre-trained on monolingual and parallel data, and the instruct model was then tuned for translation tasks only.

Send a text and the target language
The 7.8B model translates it
Term-aware or markup-safe options apply
You read the translated text

Good to know:

• Translation only: the card says it lacks chat capabilities and was trained with no chat instructions.
• Use limits: it is not meant for harming others or violating human rights, and downstream use must follow the law.
• License: GPL-3. The release uses data that requires GPL-3 distribution.
• Cost to run: none named, though a 7.8B model needs a capable machine.
• Data it reaches: it downloads files from Hugging Face, and the source names nothing else.
• Not for: chat, or production use without a proper risk assessment.

On Hugging Face, 13 November 2026

• Task: translation
• Library: transformers
• Parameters: 7.8B
• Weights: about 15.5 GB in BF16; about 3.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,095
• Likes: 30
• License: gpl-3.0
• Languages: bg, ca, cs, cy, da, de, el, en
• Based on: BSC-LT/salamandra-7b
• Quantized versions: 8
• Fine-tunes: 2
• Tags: transformers, safetensors, llama, text-generation, translation

Links

• Model: https://huggingface.co/BSC-LT/salamandraTA-7b-instruct
• Files: https://huggingface.co/BSC-LT/salamandraTA-7b-instruct/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BSC-LT/salamandraTA-7b-instruct

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