India is making a major move to build its own sovereign AI model, one that can stand alongside big global AI giants like ChatGPT (from the US-based OpenAI) and DeepSeek (from China). The government has selected Bengaluru-based Sarvam AI to lead the development of the nation’s first sovereign large language model (LLM) under the IndiaAI Mission.
Notably, this initiative aims to establish an indigenous AI model capable of advanced reasoning, speech processing, and fluency in Indian languages. To facilitate this project, Sarvam AI will be provided with access to 4,086 NVIDIA H100 GPUs over a six-month period, enabling the startup to build the LLM from the ground up.
The development will contain three distinct variants – ‘Sarvam-Large’ for complex reasoning and generation tasks, ‘Sarvam-Small’ for real-time interactive applications, and ‘Sarvam-Edge’ for on-device operations. In collaboration with AI4Bharat (IIT Madras), Sarvam AI aims to ensure that the models are deeply rooted in Indian linguistic and cultural contexts.
It’s official 🇮🇳
We’re proud to announce that Sarvam has been selected by the Government of India under the IndiaAI Mission to build India’s sovereign Large Language Model.
Building India’s sovereign model from the ground up is a crucial step toward Atmanirbhar Bharat. The… pic.twitter.com/nA4WgQ0aw0
— Sarvam AI (@SarvamAI) April 26, 2025
Sarvam AI has already proven to be among the leading AI upstarts, not just within India, but globally as well, specially when it comes to AI across languages. For example, in October 2024, the company launched Sarvam-1, a 2-billion-parameter LLM specifically optimized for Indian languages. This model supports ten major Indian languages, including Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu (along with English). In fact, unlike many existing models that struggle with token inefficiency in Indic scripts, Sarvam-1 achieves fertility rates of 1.4 to 2.1 tokens per word, significantly improving processing efficiency.
The model was trained entirely within India, using domestic AI infrastructure powered by NVIDIA H100 Tensor Core GPUs, Yotta’s data centers, and AI4Bharat’s language resources. Even, according to the company, performance benchmarks indicate that Sarvam-1 not only matches but, in some cases, surpasses larger models like Meta’s Llama 3.1 8B and Google’s Gemma-2-9B, particularly in tasks involving Indic languages. It achieved an accuracy of 86.11 on the TriviaQA benchmark across Indic languages, outperforming Llama-3.1 8B’s score of 61.47.
However, despite having proven its capabilities with Sarvam-1, building the first indigenous foundation model is still not going to be an easy task. One major hurdle is scaling up infrastructure. Training large models requires access to massive computational power over extended periods. Even with the government providing thousands of NVIDIA H100 GPUs; managing, optimizing, and maintaining such high-end resources is a complex task.
Additionally, curating high-quality, diverse datasets will be another critical challenge for the company. India’s linguistic landscape is extremely complicated, with variations not just between languages but also within dialects, cultures, and writing styles. Therefore, creating a balanced dataset that truly captures this diversity (without introducing biases) is essential, but extremely challenging.
Content originally published on The Tech Media – Global technology news, latest gadget news and breaking tech news.
