🇲🇳 Argos Translate Mongolian Models (v1.1.0)
High-speed, offline neural translation packages (.argosmodel) for Argos Translate and LibreTranslate, specifically optimized for Mongolian ↔ English translation.
🚀 What's New in v1.1.0 (October 20, 2026)
- Updated
en-mn.argosmodel: Replaced with the clean deterministic subword checkpoint (Step 5). - Zero
<unk>Glitches: Eliminates subword sampling noise and broken tokens on standard queries. - Fixed IT & Conversational Terminology:
- "Bug fixing" ➔ Алдаа засах (fixed package mistranslation).
- "Database migration completed" ➔ Өгөгдлийн сангийн шилжилт бүрэн дууссан.
- Proper future tense handling ("I have a meeting tomorrow" ➔ Би маргааш уулзалттай байна).
📦 Packages Available
| Package | Direction | Size | Quantization | Recommended Use |
|---|---|---|---|---|
en-mn.argosmodel |
English ➔ Mongolian | 137 MB | Float16 / CTranslate2 | General UI, short text & offline bots |
mn-en.argosmodel |
Mongolian ➔ English | 155 MB | Float16 / CTranslate2 | General conversational & IT translation |
🛠️ Installation & Usage
1. Using Argos Translate CLI
# Install the models
argos-translate-cli install-package en-mn.argosmodel
argos-translate-cli install-package mn-en.argosmodel
# Translate English to Mongolian
argos-translate-cli translate --from en --to mn "Where is the hotel?"
# Output: Зочид буудал хаана байна?
# Translate Mongolian to English
argos-translate-cli translate --from mn --to en "Би програм хангамжийн инженерээр ажилладаг"
# Output: I work as a software engineer
2. Python API
import argostranslate.package
import argostranslate.translate
# Install packages
argostranslate.package.install_from_path("en-mn.argosmodel")
# Translate
installed_languages = argostranslate.translate.get_installed_languages()
from_lang = list(filter(lambda x: x.code == "en", installed_languages))[0]
to_lang = list(filter(lambda x: x.code == "mn", installed_languages))[0]
translation = from_lang.get_translation(to_lang)
print(translation.translate("Good morning."))
# Өглөөний мэнд.
🧠 Architecture & Training
- Framework: OpenNMT-py + SentencePiece tokenization.
- Inference Engine: CTranslate2 (Float16 quantized).
- Latency: <25ms per sentence on standard CPU hardware.
- License: Apache 2.0.
