Sarvam AI, a cutting-edge AI tech startup from India, introduces its multilingually trained Large Language Model (LLM), Sarvam-M; making India an AI tech product market economy, powered by the first-of-its-kind MadeIn-India multilingual #LLM aka ‘Sarvam-M’.
This 24-billion-parameter open-weighs model was developed based on the Mistral AI “Small” model and forms a major step forward for India’s aspirations in the AI space.
Sarvam-M supports ten large Indian languages, Hindi, Bengali, Gujarati, Kannada, Malayalam, among others, and English, making it suitable for a variety of applications.
The newly-released LLM features an integrated thinking model which enables it to respond to not only general conversational tasks, but also more complicated logical reasoning such as mathematics and programming.
This feature makes Sarvam-M a valuable resource in the development of advanced conversational agents, advanced translation services, and innovative educational tools for Indian languages. The model is particularly strong for its Indic-language pre-training, and can be further improved post-training, thus better aligning it to the multi-lingual nature of India.
Sarvam AI was the first company in India selected by global government under the IndiaAI mission to build a foundational model.
Startup was granted access to 4,096 Nvidia H100 GPUs for Sarvam-M training for 6 months, underlining its national significance. With $53.6 million in funding and a $111 million valuation, Sarvam AI is trying to bring generative AI to every corner of India.
“I don’t think all my topnotch cast would have been ready to chip in this much with their full heart and soul and with a lot of conviction to be part of Sarvam-M,” says director Jacob, who the industry believes is an intelligent film maker.
The first two days the model had a little over 300 downloads. By May 27, more than 1,200 downloads had occurred via Hugging Face. Industry, however, is in debate about India’s preparedness and strategies to win the global AI race that is characterised by intense competition.
According to Deedy Das, a venture capitalist, he was disappointed with the early traction, contrasting it to a college project in Korea that had similar numbers of downloads in a much smaller window.
This led to a broader conversation on social media about the present state of India’s AI prowess and about what expectations there should be regarding indigenous AI models.
But There were a few voices from the tech community rallying in favor of Sarvam AI. Sridhar Vembu, Founder, Zoho reiterated that early traction does not imply long-term success and that the SarvaM team should not give up hope. Even successful products sometimes take a long time to see that initial trial period, he said.
Sarvam AI perceives Sarvam-M as the initial in a long line of contributions that would go a long way in establishing a sustainable sovereign AI ecosystem in India.
The company intends to introduce additional models and will provide comprehensive technical findings in the coming weeks. Sarvam-M is now available via Sarvam’s API and Hugging Face to allow developers and researchers to discover and contribute into its evolution.
It may have a difficult time overcoming the initial skepticism and reaction to Sarvam-M, but for India’s AI ambitions, there is no denying that this is a major step forward. The model’s specialization on Indian languages, as well its hybrid reasoning capability, cater to specific requirements within the country.
Whether it succeeds or fails in the long run may be one of the defining barometers of whether India can hope to have a place in the new global AI firmament. The Fate of Sarvam-M will be observed keenly by the Indian tech industry and the global AI one alike.