Hy-MT2-1.8B

Hy-MT2-1.8B

Translates text among 33 languages and follows translation instructions, in a 1.8B model from Tencent. For developers and teams who want a small translation model they can run themselves.

Made by tencent 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: Hy-MT2-1.8B is the smallest model in Tencent's Hy-MT2 family of multilingual translation models. The family spans 1.8B, 7B and 30B-A3B sizes. All of them translate among 33 languages and follow translation instructions written in several languages.

Who it suits: Developers, localization teams and researchers who want an open translation model they can run on their own hardware. The 1.8B size suits setups with limited memory, and quantized copies go smaller still.

Industry | Who uses it | How they use it | Example
Localization | Localization teams | Translate product text and help pages between languages | Turn a help article from English into Spanish
Software | Developers of on-device apps | Run a small translation model on the user's own machine | Use a quantized copy for translation without a cloud service
Media | Subtitle and caption editors | Draft translated text for people to review | Translate a caption file line by line, then edit it
Research | Machine translation researchers | Compare how well models follow translation instructions | Test the model with the IFMTBench benchmark released alongside it

You can skip it if you need a language outside its 33, or a hosted service with no setup.

Where to use it:

• Windows, macOS and Linux with Python, through the transformers library
• llama.cpp, using the GGUF copies listed on the card
• On-device setups, using the smaller quantized builds
• Agent tools, through the Hy-MT2-Translator Skill the card mentions on ClawHub and SkillHub

Use case | Who | What they get back
Text translation | Writers and teams | The text rendered in another of the 33 languages
Instructed translation | Developers | A translation that follows an instruction such as a tone or format
On-device translation | App builders | Translation from a small local build with no cloud call

How to set it up: You need Python and disk space for the weights.

• Python on Windows, macOS or Linux
• The transformers library, which the page lists for this model
• About 4.1 GB for BF16 weights, or about 1 GB at 4-bit as an estimate

1. Install Python if you do not have it.
2. Install the Hugging Face hub tool.
3. Download the model files.
4. Load the model with transformers, following the model card.

pip install -U huggingface_hub
hf download tencent/Hy-MT2-1.8B

How to use it: You give the model text and tell it which language to translate into. The card says it follows such instructions in several languages. The part of the card shown here has no run script, so take the loading code from the full card.

The family also comes in other builds, which the card lists in a table.

Tool | What it does
Hy-MT2-1.8B | The base 1.8B translation model
Hy-MT2-1.8B-FP8 | The 1.8B model with FP8 quantization
Hy-MT2-1.8B-GGUF | The 1.8B model for llama.cpp
Hy-MT2-1.8B-2bit-GGUF | The llama.cpp build at 2 bit
Hy-MT2-1.8B-1.25bit-GGUF | The llama.cpp build at 1.25 bit
Hy-MT2-7B | The larger 7B translation model
Hy-MT2-30B-A3B | The largest model, a mixture of experts

Example: You ask it to translate a short paragraph from Chinese into English. It returns the English text. You can add an instruction about the style you want, since the model is built to follow them.

For newcomers and experts: This is a translation engine you download and run yourself rather than a website. Start by downloading the files, then translate one sentence between two languages you can check, such as English and Chinese.

The card calls the family fast-thinking models built for difficult translation situations. For on-device use, the card says AngelSlim 1.25-bit quantization cuts the 1.8B model to 440 MB and speeds up inference by 1.5 times. It also reports that the 1.8B model beats mainstream commercial translation APIs overall, a claim to check on your own text.

Download the model files
Load the model with transformers
Pass text and a target language
Read the translation it returns

Title: Weight size in GB
BF16 weights: 4.1 GB
4-bit weights (estimate): 1 GB

Good to know:

• It has 2B parameters in Safetensors format, and the page lists a transformers library tag.
• Other people have published 57 quantized and 15 fine-tuned versions.
• License: apache-2.0.
• Cost to run: none named, you run it on your own hardware.
• Data it reaches: the files you download from Hugging Face and the text you give it to translate.
• Not for: the card names no limits beyond its 33 languages.

On Hugging Face, 13 November 2026

• Task: translation
• Library: transformers
• Parameters: 2B
• Weights: about 4.1 GB in BF16; about 1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 22,844
• Likes: 1,285
• License: apache-2.0
• Languages: zh, en, fr, pt, es, ja, tr, ru
• Quantized versions: 57
• Fine-tunes: 16
• Tags: transformers, safetensors, hunyuan_v1_dense, text-generation, translation

Links

• Model: https://huggingface.co/tencent/Hy-MT2-1.8B
• Files: https://huggingface.co/tencent/Hy-MT2-1.8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/Hy-MT2-1.8B

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