Hy-MT2 Pinokio

Hy-MT2 — A Gradio web interface for Tencent's Hy-MT2 translation models, packaged for Pinokio.

Repository: Tencent-HY-MT2-Pinokio

Overview

Hy-MT2 is a family of fast-thinking multilingual translation models. This launcher supports:

• Hy-MT2-1.8B: Lightweight model for edge devices and real-time translation
• Hy-MT2-7B: Higher-accuracy model for complex translation tasks
• Hy-MT2-30B-A3B: MoE flagship model (30B total, 3B active per token) for best quality

All three models support multilingual translation; the interface offers 38 language and variant choices with instruction-following modes such as terminology, style, contextual (background), and delimiter preservation.

Features

• 38 language and variant choices: including Traditional Chinese and Cantonese
• Translation Modes (all seven official Hy-MT2 task types):
  • Basic: Default translation
  • Terminology: Translation with custom terminology guide
  • Style: Translation with a target style (formal, casual, etc.)
  • Personalization: Translation with numbered user preferences
  • Delimiters: Preserve delimiter symbols in the output
  • Structured Data: Translate user-facing text in JSON/YAML/XML/etc. while preserving structure
  • Contextual: Translation with background information
• Model Selection: Choose between 1.8B (fastest), 7B (balanced), or 30B-A3B MoE (best quality)
• Customizable Parameters: Temperature, top-p, top-k, and repetition penalty
• Web Interface: Gradio UI accessible via browser

Installation

Using Pinokio

1. Install the app through Pinokio
2. Click Start to launch the Gradio interface
3. Open the web UI from the Pinokio interface

Manual Installation

1. Clone or download this repository

2. Install dependencies (from the app folder):

cd app
uv venv env
uv pip install --python env/Scripts/python.exe torch==2.7.0 -r requirements.txt

The command above is for Windows. On Linux/macOS, use env/bin/python instead of env/Scripts/python.exe. For GPU support, install the appropriate PyTorch build for your hardware first, as shown in PyTorch installation instructions. The Pinokio installer selects its backend automatically.

3. Run the interface from the same app folder:

env/Scripts/python.exe app.py

On Linux/macOS, run env/bin/python app.py. Open the local URL printed in the terminal (the first available port starting at 7860).

Usage

Basic Translation

1. Select source language and target language
2. Choose a model (1.8B, 7B, or 30B-A3B)
3. Enter text in Source Text
4. Click Translate

Terminology Mode

1. Select terminology from Translation Mode
2. Enter terminology guide, e.g. AI -> 人工智能 (one pair per line; lines in any other format are rejected)
3. Enter your text and click Translate

Style Mode

1. Select style from Translation Mode
2. Enter a target style, e.g. formal or literary
3. Enter your text and click Translate

Contextual Mode

1. Select contextual from Translation Mode
2. Enter background information that helps disambiguate the source text
3. Enter your text and click Translate

Personalization Mode

1. Select personalization from Translation Mode
2. Enter one preference per line (e.g. Use concise wording)
3. Enter your text and click Translate

Structured Data Mode

1. Select structured_data from Translation Mode
2. Choose the format type (JSON, YAML, XML, etc.)
3. Paste structured content in Source Text
4. Click Translate

Delimiters Mode

1. Select delimiters from Translation Mode
2. Enter text containing delimiter symbols to preserve
3. Click Translate

Supported Languages

Model Links

• Hy-MT2-1.8B: Hugging Face
• Hy-MT2-7B: Hugging Face
• Hy-MT2-30B-A3B: Hugging Face

Models are downloaded automatically from Hugging Face on first use.

Generation Parameters

Recommended parameters (pre-set in the interface; sliders auto-update when you change model):

Hy-MT2-1.8B / Hy-MT2-7B

• Temperature: 0.7
• Top-p: 0.6
• Top-k: 20
• Repetition Penalty: 1.05
• Max tokens: 4096

Hy-MT2-30B-A3B (MoE)

• Temperature: 0.7
• Top-p: 1.0
• Top-k: -1 (disabled)
• Repetition Penalty: 1.0
• Max tokens: 4096

Requirements

• Python 3.10+
• PyTorch 2.7.0 (installed by the launcher)
• CUDA-capable GPU (recommended)
  • Hy-MT2-1.8B: ~4 GB VRAM (BF16)
  • Hy-MT2-7B: ~16 GB VRAM (BF16)
  • Hy-MT2-30B-A3B: ~60 GB for BF16 weights alone, plus runtime memory; fewer active parameters do not reduce stored weights. CPU offload needs sufficient system RAM
• Transformers 5.6.0+
• Gradio 5.50.0
• Windows AMD uses CPU inference; DirectML is not integrated. Intel macOS is unsupported by the required PyTorch version.

Command Line Options

cd app
python app.py --help

Options:

• --share: Create a public Gradio link
• --server-name: Server hostname (default: 127.0.0.1)
• --server-port: Server port (default: first available starting at 7860)

Pinokio Commands

• Install: Sets up the Python environment and installs dependencies
• Start: Launches the Gradio web interface
• Update: Pulls the latest launcher changes with a fast-forward merge and reruns dependency installation
• Reset: Removes the virtual environment
• Save Disk Space: Deduplicates redundant library files

API

Use the named /translate endpoint on the URL printed at startup. The following examples assume port 7860. Inputs use the exact language labels shown in the UI, in the order below; output is [translation_text, status_message].

Python (install gradio_client):

from gradio_client import Client

client = Client("http://127.0.0.1:7860")
translation, status = client.predict(
    "Hello, how are you?", "英语 (English)", "中文 (Chinese)",
    "tencent/Hy-MT2-1.8B", "basic",
    "", "", "", "",  # terminology, context, target_style, preferences
    "JSON", 0.7, 0.6, 20, 1.05,
    api_name="/translate",
)
print(translation)

JavaScript (install @gradio/client):

import { Client } from "@gradio/client";

const client = await Client.connect("http://127.0.0.1:7860");
const result = await client.predict("/translate", [
  "Hello, how are you?", "英语 (English)", "中文 (Chinese)",
  "tencent/Hy-MT2-1.8B", "basic",
  "", "", "", "", // terminology, context, target_style, preferences
  "JSON", 0.7, 0.6, 20, 1.05,
]);
const [translation, status] = result.data;
console.log(translation);

Curl (Bash syntax; use curl.exe and adapt quoting in PowerShell):

curl -X POST http://127.0.0.1:7860/gradio_api/call/translate \
  -H "Content-Type: application/json" \
  -d '{"data":["Hello, how are you?","英语 (English)","中文 (Chinese)","tencent/Hy-MT2-1.8B","basic","","","","","JSON",0.7,0.6,20,1.05]}'

Copy the returned event_id, then retrieve the event stream:

curl -N http://127.0.0.1:7860/gradio_api/call/translate/EVENT_ID

The complete event contains the two output values. See the app's Use via API footer link for its live schema and the Gradio curl guide for event handling.

Development checks

python -m unittest discover -s app -p test_app.py -v
node --test tests/launchers.test.js

These regression tests use mocked model dependencies; they do not download weights or validate GPU translation quality.

With Gradio 5.50.0 installed, run python tests/smoke_gradio.py to check interface construction, the API schema, curl routes, and a blank request against real Gradio.

Notes

• First translation may take longer while the model downloads and loads
• GPU is recommended for faster inference
• The 1.8B model is fastest; the 7B model is more accurate for complex text
• The 30B-A3B MoE model offers the best quality but requires substantially more GPU memory
• Prompt templates follow the official Hy-MT2 documentation

License

Apache 2.0 — see the model card on Hugging Face.

References

• Hy-MT2-1.8B on Hugging Face
• Hy-MT2 Collection
• Hy-MT2 Report (arXiv:2605.22064)

Contact

For questions about the Hy-MT2 models: [email protected]