Advancing multimodal and agentic AI: systems, storage & scalability

Talk Overview
Since its inception, Sarvam has been pioneering work on Speech and Text models in Indian languages, and now we are leading the efforts on developing foundational models for India. We have released 2 major models: Sarvam-1, which is a 2B parameter model, delivering SOTA performance on the top 10 Indian languages and English. It is also able to give comparable performance to bigger 7B and 8B models, while requiring lower compute and costs for inference. Details: https://huggingface.co/sarvamai/sarvam-1

Another model is Shuka, which is an audio-native language model which accepts speech inputs and gives text output. The model is able to deliver SOTA performance on 10 Indian languages + English, and is already in the ranks of Gemini Flash, Llama3 and GPT-40-mini on Indian speech datasets, while only being an 8B parameter model. It was recently featured by Meta AI as a part of its globel GenAI success stories blog: https://ai.meta.com/blog/sarvam-india-audio-language-model-llama/

Key takeaways

1. Spread awareness about the innovation in Indian foundational models
2. Encourage develop community to build on top of these foundational models and how Sarvam can support them

Target Audience

1. Developers building applications on top of speech APIs and foundational models
2. Enterprises building for Indian customers and looking for Voice AI solutions

Speaker Bio:
Tushar Goswamy: I am working on GTM & Strategy at Sarvam, and leading some of our new verticals like Edge AI. I am also leading partnerships for Sarvam with leading companies like Meta & Google as well as with select government departments like Ministry of Skill Development. I was previously working as a Sr. Consultant with McKinsey, where I led new business development projects in sectors like Electric Vehicles, Drones & GenAI. I did my undergrad from IIT Kanpur, where I co-founded 2 AI startups and dabbled with AI research, having published at conferences like NeurIPS, ICML and COLING.