🇮🇳 BharatBhasha AI - Machine Translation (T2T) Models

Production-grade, INT8-quantized ONNX neural machine translation packages for AI4Bharat IndicTrans2, specifically optimized for offline deployment on Android mobile devices and edge environments.

📦 Model Artifacts & Architecture

Model Bundle	Direction	Languages Supported	Size	Format
`indictrans2-en-indic-dist-200M-ONNX-int8`	English → Indic	22 Scheduled Indian Languages	~235 MB	INT8 ONNX
`indictrans2-indic-en-dist-200M-ONNX-int8`	Indic → English	22 Scheduled Indian Languages	~235 MB	INT8 ONNX
`indictrans2-indic-indic-dist-320M-ONNX-int8`	Indic → Indic	Any pair among 22 Languages	~325 MB	INT8 ONNX

Bundle Contents

Each model directory includes:

encoder_model.onnx: Quantized INT8 transformer encoder
decoder_model.onnx: Initial token decoder
decoder_with_past_model.onnx: Optimized decoder with past KV-cache reuse for low-latency generation
dict.SRC.json, dict.TGT.json: Vocabulary maps
model.SRC, model.TGT: SentencePiece tokenizer models
tokenizer_src.json, tokenizer_tgt.json: HuggingFace Fast Tokenizers

📱 Mobile Android Deployment (ONNX Runtime Mobile)

Add to build.gradle.kts:

dependencies {
    implementation("com.microsoft.onnxruntime:onnxruntime-android:1.20.0")
}

Load the session with hardware acceleration:

val env = OrtEnvironment.getEnvironment()
val sessionOpts = OrtSession.SessionOptions().apply {
    addNnapi() // Android Neural Networks API acceleration
    setIntraOpNumThreads(4)
}
val encoderSession = env.createSession(encoderFile.absolutePath, sessionOpts)

⚡ Inference Benchmarks

Encoder Latency (Mobile CPU/NNAPI): ~85ms
Decoder Latency Per Token: ~28ms
RAM Footprint: ~160 MB (Comfortably runs on 3GB–4GB budget Android phones)

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