Daily Current Affairs 26-February-2026

At the AI Impact Summit, Bengaluru-based startup Sarvam AI unveiled two Large Language Models (LLMs) trained in India.

How are LLMs Trained?

• Large Language Models (LLMs) are trained on clusters of Graphics Processing Units (GPUs), which are specialized chips designed for parallel processing.
• Training requires thousands of GPUs working together in data centres, often consuming enormous amounts of electricity and cooling resources.
• The combined expense of hardware, electricity, and maintenance runs into millions of dollars.
• LLMs are trained on massive datasets scraped from the Internet, books, articles, and other sources.
• However, Indian languages are underrepresented compared to English, European, and East Asian languages.

Importance of India in AI & LLM Indigenisation

• Linguistic and Cultural Representation: India has 22 official languages and hundreds of dialects, which global models often fail to capture.

• Indigenous LLMs ensure better accuracy and inclusivity for Indian languages, making AI accessible to a wider population.

• Strategic and Economic Independence: It also builds domestic capacity, creating opportunities for startups, researchers, and industries to innovate locally.
• Cost Efficiency and Ecosystem Growth: Indigenous models foster a self-reliant AI ecosystem, encouraging innovation in sectors such as education, healthcare, agriculture, and governance.

Challenges in Training LLMs in India

• Data Scarcity: Limited availability of high-quality datasets in Indian languages. Models often rely on translation into English, which increases token usage and reduces efficiency.
• Capital Constraints: Training LLMs demands massive investment, often beyond the reach of Indian startups. Without immediate business use cases, raising funds becomes difficult.
• Performance in Local Contexts: Suboptimal results in Indian languages affect adoption, especially for low-cost devices like feature phones.

Government Support – IndiaAI Mission

• Commissioned 36,000 GPUs in Indian data centres (e.g., Yotta).
• Subsidised access for startups and researchers at nominal fees.
• Example: Sarvam AI received 4,096 GPUs, with subsidies worth nearly ₹100 crore.
• Aim: Build domestic AI ecosystem and reduce dependence on foreign-developed LLMs.

Mixture of Experts (MoE) Architecture

• Innovative Design: MoE is a breakthrough approach that makes Large Language Models (LLMs) more cost-efficient by reducing the computational load during inference.
• Selective Parameter Activation: Unlike traditional models where all parameters are activated for every query, MoE activates only a small subset of parameters.
• Practical Advantages: Delivers quicker response times since fewer parameters are processed. Requires less computing power and energy, lowering operational costs.
• Strategic Value: By lowering costs, MoE enables startups and research institutions to train competitive models locally.

Significance

• Sarvam’s models show progress in Indian language accuracy and efficiency.
• Other firms like BharatGen and Gnani.ai are also building sector-specific models (education, healthcare, speech).
• India’s push mirrors global trends (e.g., China’s DeepSeek R1) in cost-effective AI innovation.

Conclusion

Training LLMs in India faces hurdles of data, capital, and infrastructure, but government support through the IndiaAI Mission and innovations like MoE architecture are helping bridge the gap. These efforts are crucial for building AI models aligned with India’s linguistic diversity and strategic needs.