Sarvam Selected to Build India’s First Indigenous AI Large Language Model
In a significant move for India’s artificial intelligence (AI) landscape, the government has selected Bengaluru-based start-up Sarvam to develop the country’s first indigenous AI large language model (LLM). This decision comes amid growing competition from China’s DeepSeek, a low-cost AI model that has made waves in the global market. Sarvam was chosen from a pool of 67 applicants and will receive vital government support in the form of compute resources to build the model from the ground up.
India’s Bold AI Initiative with Sarvam
Sarvam is the first start-up to be approved for financial and infrastructure assistance under the IndiaAI Mission, a government-backed program with an investment of Rs 10,370 crore. The mission aims to speed up AI innovation in India and make the nation a leader in AI technologies. The government is currently scrutinizing hundreds of other proposals, but Sarvam's project has been chosen as a showcase example of India's AI dreams.
The start-up’s AI model is designed to be capable of advanced reasoning and fluent in multiple Indian languages, aiming for large-scale deployment. Sarvam’s CEO, Pratyush Kumar, explained that the model would be optimized for voice, making it ideal for a wide range of applications in India.
Also read: Huawei to Test New Ascend 910D AI Processor
Government Support and Cutting-Edge Resources
Sarvam’s AI project will receive significant backing from the government, including access to 4,000 graphics processing units (GPUs) for six months to build and train the model. These GPUs will be supplied by companies chosen by the government to set up AI data centres across India. While the model won’t be open-sourced, it will be fine-tuned specifically for Indian languages and use cases, setting it apart from other models in the market.
Ashwini Vaishnaw, India’s IT Minister, highlighted that the model would feature 70 billion parameters, with several innovations in both engineering and programming. According to Vaishnaw, this would allow Sarvam’s model to compete with the best AI models in the world, solidifying India’s position in the global AI race.
Three Model Variants for Different Applications
Sarvam’s LLM proposal includes three distinct variants: Sarvam-Large, Sarvam-Small, and Sarvam-Edge. Sarvam-Large is designed for advanced reasoning and generation, Sarvam-Small will focus on real-time interactive applications, and Sarvam-Edge will be aimed at compact, on-device tasks. This range of models ensures the AI can be applied across multiple use cases, from high-level decision-making to more localized, immediate tasks.
The development of this AI model comes amid the rapid rise of DeepSeek, China’s foundational AI model. DeepSeek has disrupted the market with its low cost, open-source nature, and claimed superior accuracy despite being trained on lower-grade GPUs compared to models from companies like OpenAI. This development has sent shockwaves through the AI industry, affecting companies like Nvidia, which saw its stock decline following DeepSeek’s entry.
India’s Vision for AI Autonomy
Sarvam’s AI model will be built, deployed, and optimized entirely in India, using local infrastructure and talent. The initiative is not just about technology but also about promoting strategic autonomy for India in the AI domain. Sarvam envisions a future where India leads the global AI charge, driving innovation from within the country while ensuring that data and applications stay within national borders.
Vivek Raghavan, co-founder of Sarvam, emphasized that this project is a key step in building critical national AI infrastructure. He noted that once these foundation models are created, they will unlock a universe of possibilities for both citizens and enterprises. For individuals, it means interacting with AI that is culturally relevant and easy to use. For businesses, it means leveraging intelligence without the need to send data abroad.
Earlier this year, the government had also selected 10 companies to provide 18,693 high-end GPUs for developing machine learning tools. This is significantly more than the initial target of 10,000 GPUs set under the IndiaAI Mission, highlighting the growing ambition to develop foundational AI models that will position India as a global leader.
