How BharatGen is building India AI with NVIDIA Nemotron

The seeds were sown through a consortium of academic institutions. IIT Bombay took the lead. We went through a call for proposals and we were selected to build the small models up to 10 billion parameters with 235 crores funding. Over a period of two years, we applied to the India IIM mission for building models up to 1 trillion parameters. The proposal was looked at very favorably and the whole initiative grew into a completely government-owned section. We just had a whole bunch of models launched with the honour of the PM, multilingual, across 12 languages, the speech models, and 22 languages, text models. And we're very happy with the collaborations that have also grown around this, with hospitals, with state governments, with several line departments and departments of water and sanitation. How can we use our workload with healthcare, making it easily comprehensible by the patient, making it possible for an ordinary person to learn other languages and communicate in that language. There are so many interesting use cases with Gyan Bharatam, for example. There have been lots of interesting engagements in the finance sector as well. And even in the government space, education is a huge potential. The work we've been doing with Kotak Education Foundation and Spoken Tutorials, giving them both co-pilots that will help them converse better and be more effective in their assessments. So we use them as catalysts, as very good learning points. NVIDIA has been a solid partner in that entire journey. We went through many cycles. We started with Megatron, then Nemotron. We did find a very healthy ecosystem building around Nemotron. So we're grateful for that, especially on the pre-training and post-training part. You know, open source, which is what NVIDIA stands for, right? Open source is not just the code, but is also the dialogue. We're very grateful to NVIDIA for helping us scale our model to multiple nodes. But again, it's not just scaling to multiple nodes. How is it being scaled? How do you maintain good efficiency and throughput? So observability comes very handy there. The development flexibility and scalability has been very important learning for us in our journey together with NVIDIA. The BCM and SLURM by LCP, NEMO 2.0 for the pre-training, NEMO RL for the post-training, are very specific examples of how we have leveraged the open source ecosystem that NVIDIA has been creating. And while that is very important in terms of training, there's also another important aspect, which is inference. We are in conversation with NVIDIA on how we could do the inference. We'll continue working together, build large models. Anyone who stands for open source is a partner for life. We like to work with NVIDIA's open source initiators. And as I said, it triggers both in-house capability, avoids reinvention of the wheel, and I think it contributes to the community as a whole.