Analyzing Quantization in Multilingual Transformer Models for Indic Translation

The growing adoption of multilingual sequence-to-sequence transformer models has significantly advanced neural machine translation (NMT), enabling support for hundreds of language pairs. However, their substantial computational and memory requirements hinder deployment in resource sparse environments. Post-Training Quantization (PTQ) addresses this by reducing model size and inference cost through low-precision representations. While PTQ has been studied for high-resource, Latin-script languages, its effects on translation tasks involving Indic languages remain underexplored. This work investigates how quantization impacts translation quality in morphologically rich, low-resource, non-Latin script languages, using Hindi as a representative case. This model evaluates three multilingual transformer models, mBART, M2M-100, and NLLB, under four quantization levels (FP32, FP16, INT8, and 4-bit) across three language pairs: English-French (high-resource Latin script), English-Galician (low-resource Latin script), and English-Hindi (low-resource Indic), using the FLORES-101 benchmark. Results show that Hindi often exhibits degradation levels comparable to or lower than Galician and even French. FP16 and INT8 quantization maintain strong translation quality, while 4-bit variants degrade significantly. Reverse translations (target language to English) consistently outperform forward translations. These findings suggest that moderate quantization is viable for Indic-language translation and that degradation is not solely dictated by language resource level or script. This work provides systematic observations of quantized multilingual translation for Indic languages, offering a foundation for future research in low-resource, non-Latin script NMT.