Hy-MT2-1.8B-1.25Bit-GGUF

Hy-MT2-1.8B-1.25Bit-GGUF

Translates between 33 languages on your own device with a 1.8B model squeezed to 1.25 bits, about 440 MB as GGUF files for llama.cpp. For developers adding offline translation to apps.

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: This is a 1.25-bit GGUF build of Hy-MT2-1.8B, a Tencent translation model that covers 33 languages. AngelSlim did the quantization. The card says it shrinks the model to 440 MB and speeds up inference by 1.5 times, which makes it a fit for on-device use.

Who it suits: It suits developers who want translation running locally, in an app or on a small machine, without sending text to a server.

Industry | Who uses it | How they use it | Example
Apps | Mobile and desktop developers | Add offline translation to an app | Translating a chat message on the device
Localization | Localization teams | Draft translations between the supported languages | Turning a product string into another language
Edge devices | Embedded developers | Run translation where storage is tight | Fitting a model of about 440 MB
Research | Translation researchers | Compare quantization levels of one model | Choosing between the 1.25-bit, 2-bit and full files

If you want a larger model, you can skip this one, since the family also has 7B and 30B-A3B sizes.

Where to use it:

• llama.cpp, which the card names for its GGUF builds
• Windows, macOS or Linux, with Python for downloading the files
• Tencent's Hy-MT2-Translator skill, which the card offers for bringing the models into translation tasks

Use case | Who | What they get back
On-device translation | App developers | Translated text with no server call
Instruction-guided translation | Users | Translations that follow instructions written in several languages
Quantization comparison | Researchers | A tiny build to set against the 2-bit and full versions

How to set it up:

• Python and the Hugging Face hub package, to download the files
• llama.cpp, installed by its own instructions
• About 440 MB of space for the model

1. Install the Hugging Face hub package.
2. Download the model files with the command below.
3. Install llama.cpp.
4. Point llama.cpp at the downloaded GGUF file and give it a translation request.

pip install -U huggingface_hub
hf download tencent/Hy-MT2-1.8B-1.25Bit-GGUF

How to use it: The card shows no run command for this build. You load the GGUF file in llama.cpp, then ask for a translation and name the target language. The card says the model follows translation instructions written in several languages.

Title: Builds in the Hy-MT2 family
Model | What it is
Hy-MT2-1.8B | The 1.8B translation model in full form
Hy-MT2-1.8B-FP8 | The 1.8B model with FP8 quantization
Hy-MT2-1.8B-GGUF | The 1.8B model as GGUF files for llama.cpp
Hy-MT2-1.8B-2bit-GGUF | The 1.8B model as 2-bit GGUF files
Hy-MT2-1.8B-1.25bit-GGUF | The 1.8B model as 1.25-bit GGUF files
Hy-MT2-7B | The 7B translation model
Hy-MT2-7B-FP8 | The 7B model with FP8 quantization
Hy-MT2-7B-GGUF | The 7B model as GGUF files for llama.cpp
Hy-MT2-30B-A3B | The 30B-A3B translation model
Hy-MT2-30B-A3B-FP8 | The 30B-A3B model with FP8 quantization

Example: You give the model a short English sentence and ask for it in another of the 33 languages, in the instruction language you prefer. A translated sentence comes back. The card text shown here gives no sample output.

For newcomers and experts: GGUF is a file format that llama.cpp can run on an ordinary computer. A 1.25-bit file is squeezed hard, so it is small and quick. A good first try is one short sentence, checked by someone who reads both languages.

The Hy-MT2 line has three sizes, 1.8B, 7B and 30B-A3B, built to give quick translation answers. This build comes from the 1.8B model through AngelSlim quantization. The card says the full 1.8B model beats mainstream commercial translation APIs overall, but it lists no scores for this 1.25-bit file.

Text and a translation instruction
1.25-bit 1.8B model in llama.cpp
Translated text comes out

Good to know:

• Squeezing weights to 1.25 bits usually costs some quality, and the visible card gives no scores for this build.
• The release also includes IFMTBench, a benchmark for how well a model follows translation instructions.
• The 440 MB and 1.5 times figures come from the card, and they describe the 1.25-bit build.
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: none named beyond the text you give it
• Not for: none named

On Hugging Face, 13 November 2026

• Parameters: 1.8B
• Weights: about 896 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 11,029
• Likes: 50
• License: apache-2.0
• Based on: tencent/Hy-MT2-1.8B
• Tags: gguf, conversational

Links

• Model: https://huggingface.co/tencent/Hy-MT2-1.8B-1.25Bit-GGUF
• Files: https://huggingface.co/tencent/Hy-MT2-1.8B-1.25Bit-GGUF/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-1.25Bit-GGUF

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