BharatBhasha AI - Machine Translation (T2T) Models Catalog

📌 Repository Purpose & Motivation

This repository provides full-precision and quantized INT8 ONNX packages for AI4Bharat IndicTrans2, enabling offline neural machine translation across all 22 Official Scheduled Indian Languages + English.

Why this repository exists:

Un-gated Direct Distribution: Upstream Hugging Face repos may enforce gating forms, authentication restrictions, or experience availability drops. This un-gated mirror allows Android and desktop apps to pull translation weights directly without requiring user accounts or tokens.
Dedicated Translation Bandwidth: Machine translation involves multi-file encoder/decoder ONNX models with past-key-value caches and custom subword tokenizers. Isolating T2T from speech models prevents download queue blockage.
Optimized for On-Device Execution: Quantized to INT8 ONNX to fit inside mobile device memory constraints while maintaining BLEU score parity with FP32 models.

🏗️ Repository Folder Structure

The repository is organized into 3 distinct model packages, each containing the encoder, decoder, shared weights, dictionary mappings, and tokenizer configurations:

remiai3/TRANSLATION_MODELS/
├── .gitattributes
├── README.md                                          # Documentation & Catalog
│
├── indictrans2-indic-indic-dist-320M-ONNX-int8/       # Universal Indic-to-Indic (320M INT8)
│   ├── config.json
│   ├── generation_config.json
│   ├── encoder_model.onnx
│   ├── encoder_model.onnx.data
│   ├── decoder_model.onnx
│   ├── decoder_with_past_model.onnx
│   ├── decoder_shared.onnx.data
│   ├── dict.SRC.json
│   ├── dict.TGT.json
│   ├── model.SRC
│   ├── model.TGT
│   ├── tokenizer_config.json
│   ├── tokenizer_meta.json
│   ├── tokenizer_src.json
│   ├── tokenizer_tgt.json
│   ├── special_tokens_map.json
│   ├── tokenization_indictrans.py
│   └── translate.py
│
├── indictrans2-indic-en-dist-200M-ONNX-int8/          # Any Indic Language to English (200M INT8)
│   ├── config.json
│   ├── generation_config.json
│   ├── encoder_model.onnx
│   ├── encoder_model.onnx.data
│   ├── decoder_model.onnx
│   ├── decoder_with_past_model.onnx
│   ├── decoder_shared.onnx.data
│   ├── dict.SRC.json
│   ├── dict.TGT.json
│   ├── model.SRC
│   ├── model.TGT
│   ├── tokenizer_config.json
│   ├── tokenizer_meta.json
│   ├── tokenizer_src.json
│   ├── tokenizer_tgt.json
│   ├── tokenization_indictrans.py
│   └── translate.py
│
└── indictrans2-en-indic-dist-200M-ONNX-int8/          # English to any Indic Language (200M INT8)
    ├── config.json
    ├── generation_config.json
    ├── encoder_model.onnx
    ├── decoder_model.onnx
    ├── decoder_with_past_model.onnx
    ├── dict.SRC.json
    ├── dict.TGT.json
    ├── model.SRC
    ├── model.TGT
    ├── tokenizer_config.json
    ├── tokenizer_meta.json
    ├── tokenizer_src.json
    ├── tokenizer_tgt.json
    ├── special_tokens_map.json
    ├── tokenization_indictrans.py
    └── translate.py

🌐 Direct Download Endpoints for Android & Mobile Apps

Clients can download individual files directly without an access token:

Endpoint Format

https://huggingface.co/remiai3/TRANSLATION_MODELS/resolve/main/{MODEL_NAME}/{FILE_NAME}

Examples:

Download Indic-Indic Encoder:

https://huggingface.co/remiai3/TRANSLATION_MODELS/resolve/main/indictrans2-indic-indic-dist-320M-ONNX-int8/encoder_model.onnx

Download Indic-English Config:

https://huggingface.co/remiai3/TRANSLATION_MODELS/resolve/main/indictrans2-indic-en-dist-200M-ONNX-int8/config.json

Download English-Indic Tokenizer:

https://huggingface.co/remiai3/TRANSLATION_MODELS/resolve/main/indictrans2-en-indic-dist-200M-ONNX-int8/tokenizer_src.json

📊 Translation Models Breakdown

Model Name	Parameters	Quantization	Size	RAM Overhead	Primary Function
`indictrans2-indic-indic-dist-320M-ONNX-int8`	320M	INT8	370 MB	< 350 MB	Direct translation between any pair among the 22 Indian languages.
`indictrans2-indic-en-dist-200M-ONNX-int8`	200M	INT8	257 MB	< 300 MB	High-quality translation from any of the 22 Indian languages into natural English.
`indictrans2-en-indic-dist-200M-ONNX-int8`	200M	INT8	363 MB	< 300 MB	High-quality translation from English into any of the 22 scheduled Indian languages.

⚖️ Open Source Licensing & Credits

Model Architecture & Base Weights: AI4Bharat IndicTrans2 developed by AI4Bharat (IIT Madras).
ONNX Quantization & Export: Quantized with ONNX Runtime by the community (naklitechie / hari31416).
License: MIT License (Permits free commercial and non-commercial redistribution).
Acknowledgement: Sincere gratitude to the AI4Bharat research team at IIT Madras and the Bhashini Mission for creating and open-sourcing foundational translation models for Indian languages.

Downloads last month: -; Downloads are not tracked for this model. How to track