IMC 2024: Ex-Xiaomi India head Manu Kumar Jain demoed new Hindi LLM

At the India Mobile Congress (IMC) 2024, we were met with a pleasant surprise. At the Qualcomm booth, we saw a demo by former Xiaomi India head Manu Kumar Jain. He debuted his latest project called Nanda. Jain is now the CEO of G42 India, an Abu Dhabi-based AI venture.
The model was first unveiled in September 2024, however, today was the first time that the model was demonstrated live. “Just like the prime minister said in his opening speech on Day 1 of IMC that each and everyone should benefit from AI, we aim to offer equitable AI to users in India,” said Manu Jain during the demonstration.
He says, Nanda is “probably the largest Hindi AI model”. “It writes in Hinglish better than any other model would,” he adds.
We tried out the model, which was demoed on the Qualcomm cloud AI servers; Nanda’s own UI develops on the Qualcomm interface and has been modified to look a little different. However, we interacted with the model and its Hinglish responses are pretty good. It seems quite natural because of the mix of Hindi and English words it uses in a sentence — similar to how many people speak English in India. Besides Hinglish, it also understands and responds to prompts written in Hindi language and devnagri script.
We also asked Jain if we can expect a ChatGPT-like standalone app for Nanda soon. Jain said yes. However, he added that “some version of Nanda may soon be available as standalone”.
Nanda is a 13-billion parameter AI model that has been trained using around 2.13 trillion tokens from language datasets, including content in Hindi. Its launch represents a major advancement for AI in India, empowering more than 500 million Hindi speakers with access to generative AI capabilities.
In August 2023, G42 introduced JAIS, the world’s first open-source Arabic large language model (LLM). JAIS revolutionised Arabic Natural Language Processing (NLP), offering generative AI capabilities in native Arabic to over 400 million speakers worldwide. Spanning models from 590 million to 70 billion parameters, JAIS established a new benchmark for language-specific AI. G42 is now trying to do the same with Nanda for the Indian audience. The company is also working on similar models for regions with languages that remain under-represented in AI.