India’s Breakthrough in Large Language Models Development

Recent advances in artificial intelligence (AI) have seen India emerge as player in developing Large Language Models (LLMs). At the 2026 AI Impact Summit in Bengaluru, Sarvam AI revealed two new LLMs. These models, trained on 35 billion and 105 billion parameters, demonstrate improved performance in Indian languages while consuming less power and computing resources compared to global counterparts. This marks a substantial step in India’s quest for AI models tailored to local languages and needs.

Training Large Language Models in India

LLMs are trained using clusters of Graphics Processing Units (GPUs) that require vast computational power and electricity. The training costs run into millions of dollars. Most global LLMs rely on data dominated by English and East Asian languages, leading to challenges in developing models for Indian languages. Indian startups face data scarcity and limited capital, making it difficult to create efficient LLMs without relying on translation to English. Advances in machine translation help but can reduce performance and increase computational load.

Government Support and Infrastructure

The Government of India supports domestic LLM development through the IndiaAI Mission. It has commissioned over 36,000 GPUs in data centres run by Indian firms like Yotta. Researchers and startups access these resources at subsidised rates. Sarvam AI received access to 4,096 GPUs and nearly ₹100 crore in subsidies to develop their models. The Ministry of Electronics and Information Technology promotes local LLMs to enhance Indian language capabilities and encourage AI talent. This support aims to reduce dependency on foreign AI technologies.

Innovations in Model Architecture

Sarvam AI’s models utilise the Mixture of Experts (MoE) architecture. Unlike traditional models that activate all parameters during inference, MoE activates only a subset. This reduces computing costs and speeds up responses. Although Sarvam’s largest model has 105 billion parameters, it is smaller than global frontier models but is optimised for accuracy and efficiency in the Indian context. The firm plans to scale up once further investment is available. Other Indian firms like BharatGen and Gnani.ai are also developing multilingual and speech AI models using the government-supported infrastructure.

Challenges and Future Prospects

Despite progress, Indian LLMs face hurdles such as limited data diversity and capital constraints. Open sourcing remains limited, restricting external validation. Models currently offer less depth than global paid versions but focus on practical use cases like education and healthcare. Continued government backing and ecosystem development are expected to enhance India’s AI capabilities, making LLMs more accessible and relevant to Indian users.

Topics for Prelims:

Large Language Models (LLMs)

1. LLMs are AI models trained on massive datasets using GPUs.
2. Training involves high computational costs and energy consumption.
3. LLMs use parameters to understand and generate human-like text.
4. MoE architecture activates only parts of the model for efficiency.
5. Indian languages pose unique challenges due to limited data.

IndiaAI Mission and Government Initiatives

1. IndiaAI Mission subsidises GPU access for AI research.
2. Over 36,000 GPUs commissioned in Indian data centres.
3. Supports startups and researchers with affordable compute resources.
4. Encourages development of AI models in Indian languages.
5. Promotes AI ecosystem growth and talent development.

Mixture of Experts (MoE) Architecture

1. MoE activates only a fraction of model parameters during inference.
2. Reduces computational cost and speeds up model responses.
3. Enables training of large parameter models with less energy.
4. Balances model size with efficiency and accuracy.
5. Used in Sarvam AI’s models for Indian language optimisation.

Questions for UPSC:

1. Point out the challenges and opportunities in developing artificial intelligence models for multilingual societies like India.
2. Critically analyse the role of government subsidies and infrastructure in promoting indigenous AI innovation, with suitable examples from India.
3. Estimate the impact of computational power and energy consumption on sustainable AI development and suggest measures to address these concerns.
4. Underline the significance of model architectures like Mixture of Experts in making AI accessible and efficient, and discuss their potential applications in public service delivery.

Answer Hints:

1. Point out the challenges and opportunities in developing artificial intelligence models for multilingual societies like India.

1. Scarcity of high-quality, diverse training data for many Indian languages compared to English and other global languages.
2. Complexity in handling multiple scripts, dialects, and linguistic nuances across regions.
3. Machine translation advances help but may degrade model efficiency and accuracy when relying on translation as an intermediate step.
4. Opportunity to develop AI models tailored to local languages, enhancing accessibility and digital inclusion.
5. Potential to encourage innovation in education, healthcare, governance by leveraging multilingual AI solutions.
6. Challenges in capital investment and infrastructure limit large-scale indigenous model development but government support can mitigate these.

2. Critically analyse the role of government subsidies and infrastructure in promoting indigenous AI innovation, with suitable examples from India.

1. IndiaAI Mission provides subsidised access to over 36,000 GPUs, lowering entry barriers for startups and researchers.
2. Example – Sarvam AI received 4,096 GPUs and ₹100 crore subsidy, enabling development of 35B and 105B parameter LLMs.
3. Government infrastructure reduces dependency on foreign AI technologies and promotes self-reliance.
4. Encourages talent development and ecosystem growth focused on Indian languages and contexts.
5. Subsidies address capital scarcity but open-source and transparency remain limited, potentially restricting wider collaboration.
6. Such interventions create a strategic advantage by encouraging cost-efficient AI innovations tailored to local needs.

3. Estimate the impact of computational power and energy consumption on sustainable AI development and suggest measures to address these concerns.

1. Training large LLMs requires millions of dollars and consumes vast electricity, raising environmental concerns.
2. High energy use contributes to carbon footprint and operational costs, limiting sustainability.
3. Efficient architectures like Mixture of Experts reduce active parameters, lowering compute and energy needs.
4. Government support for local data centres optimises resource use and reduces reliance on energy-intensive foreign infrastructure.
5. Promoting smaller, efficient models focused on specific languages or tasks enhances sustainability.
6. Measures – investment in renewable energy-powered data centres, model pruning, quantization, and federated learning to reduce energy consumption.

4. Underline the significance of model architectures like Mixture of Experts in making AI accessible and efficient, and discuss their potential applications in public service delivery.

1. Mixture of Experts (MoE) activates only a subset of model parameters during inference, reducing computational cost and latency.
2. Enables training and deployment of large models with fewer resources, making AI accessible to startups and regions with limited infrastructure.
3. Improves efficiency without compromising accuracy, suitable for localized language tasks.
4. Potential applications include education (multilingual tutoring), healthcare (diagnosis support), and governance (regional language chatbots).
5. Facilitates deployment on low-end devices like feature phones, increasing reach in rural and underserved areas.
6. Supports scalable AI solutions tailored to diverse linguistic and cultural contexts, enhancing public service delivery.