Limited training data is one of the central challenges in building AI models for low-resource languages. There is also an asymmetry in the availability of data resources across languages. For instance, while Hindi-English parallel corpora might be abundant, Gujarati-English or Gujarati-Hindi data is limited. A recent advancement in NLP has been the training multilingual models utilizing data from multiple languages simultaneously. Such model also low-resource languages to benefit from larger corpora in high-resource languages. In some scenarios, the models can give good performance on a language even when no training data for that language is available - referred to as zero-shot cross-lingual transfer. We have also seen benefits from multilingual models in the training of the IndicTrans translation model and other models we have trained. Multilingual learning is an important area of research and there are many open questions to explore in order to achieve high levels of cross-lingual transfer. Broadly, some research directions to explore are: How do we go beyond current transfer architectures and improve transfer across languages? Better multilingual transfer for generation tasks like one to many translation How to reduce language divergences in the multiNMT architecture? What are the language-specific biases in current multilingual modeling approaches, which can be reduced to enable better multilingual transfer. A lot of the success in multilingual learning has been in the text domain. We would like to push the limits of multilingual learning in the speech modality as well as speech+text modality.