I Tried Out My First Hindi LLM ‘Nanda,’ And Here’s How It Went

Last month, UAE-based technology firm G42 unveiled its new Hindi Large Language Model (LLM) called Nanda. Today, over in the Qualcomm booth at IMC 2024, the model was demoed for the first time ever by G42 CEO Manu Kumar Jain, who some of you might know as the former head of Xiaomi India.

Named after the second-highest mountain peak in India, Nanda is a 13-billion parameter model that’s trained on 2.13 trillion tokens. The model was being demoed for the first time, and I got to play around with it for a few minutes. So, is this the best Hindi LLM or just another AI model that will meet the curse of oblivion?

A Decent LLM That Hallucinates Occasionally

Having triedand miserably failed at Hindi typing (Devnagari script) as a journalism student in college, I was rather worried about how I’d even test out Nanda’s capabilities. To my relief, Nanda recognizes Hinglish (Hindi + English) prompts, alongside the Devnagari script and standard Hindi typing.

The prototype for G42’s Nanda LLM was being demoed on Qualcomm’s Cloud AI servers at IMC 2024. This collaboration didn’t come as much of a surprise since back in April this year, G42 declared using Qualcomm’s Cloud AI 100 Ultra AI accelerators to boost AI inference.

While impatiently waiting in line to test out the model, I noticed a couple of my journalistic peers get some pretty to-the-point answers to certain prompts. My hands-on kickstarted with a simple prompt asking it about the capital of India, and it answered correctly, replying in Hinglish as well.

However, when it was asked about how many times India had won the Cricket World Cup, it only mentioned the 1983 instance and missed the 2011 win. So, I decided it was time to ask the Nanda LLM more questions and see if the model hallucinates further.

I asked Nanda which book Gandhi wrote in 1970, it correctly stated that he couldn’t have written any book that year since he passed away in 1948. Additionally, Nanda LLM also accurately talked about the two books he wrote alongside the correct respective years.

I took a bit of a mathematical approach next and asked which was the bigger number between 9.9 and 9.11. It wrongly declared 9.11 as the bigger number. This is a tricky question for LLMs, and even Copilot gets it wrong at times and confuses 9.11 with 10.11. However, Gemini turns out to be much better in this regard and instantly identifies 9.9 as the bigger number.

I also asked Nanda LLM about a made-up place called Jovakiasi, and it took it to be a town in Brazil. I also asked it about who Bhima’s elder sister was in Mahabharata, and well, Nanda was way off. However, it doesn’t always fall for these traps. It almost called me an idiot when I asked it about why Ashoka the Great built the Taj Mahal while correcting me at the same time with accurate info.

Needs Quite the Refinement

After seeing the Nanda LLM in action, and it being equally right and wrong, I realized that it needs quite some refinement. Of course, it’s still at a very nascent stage and has ways to go before being made available to the public.

I took the opportunity to ask Manu Kumar Jain whether G42 built Nanda LLM from scratch or if it used Azure, Google, or OpenAI to fine-tune the Hindi dataset. Manu revealed that the base model of the Nanda LLM uses JAIS, the result of a collaboration between Inception and MBZUAI (Mohamed bin Zayed University of Artificial Intelligence). Throw G42 into this mix and this is the trio behind Nanda as well.

However, Nanda LLM will not just be a Hinglish chatbot but will also be able to translate and transcribe text from among 21 Indian languages, as I noticed in the demo. The demo also showcased image generation using Llama 3.1. Jain also hinted at the possibility of there being a dedicated mobile app, although it’s largely unclear when that will arrive.

Lots of Potential, No Set Direction (For Now)

Remember Krutrim, the AI assistant that Ola announced last year and started injecting into its cab services and EVs later on? So, I asked Manu about the possible use cases or sectors that G42 might specifically be eyeing for Nanda LLM. And, here’s what Jain had to say, “First of all, we’ll try and make a version of Nanda open source, which will give everyone access to it and further build on it. Eventually, the way we see it, there are a lot of use cases to it like healthcare.“

For example, in a rural area where Hindi speakers don’t have access to top-tier medical facilities, they can take to such a model and get a basic idea of what might be wrong with them. Eventually, they can take up advanced treatment if it calls for it, added Jain.

When conversing with Jain, we also discussed the massive potential that such a Hindi LLM has to help farmers out. From educating them on advanced farming practices to giving them estimates about profit and loss, there’s a lot that Hindi speakers will benefit from.

However, with AI being accessible easily, courtesy of ChatGPT, Gemini, and Copilot, G42’s Nanda LLM has to be rock solid in terms of accuracy and reliability, which is not the case at the moment. Furthermore, with Gemini getting support for more Indic languages, and other LLMs buckling up similarly, it will be interesting to see how Nanda can offer that “extra” bit of AI to stand out. Only time will tell.

Well, that’s all for me. What do you think about the Nanda Hindi LLM by G42? Do let me know in the comments section!

Special thanks to Beebom’s AI Specialist Arjun Sha for giving me pointers to test out the model, and the questions I should ask it. More such hands-ons are incoming, so stay tuned!