3. Sovereign AI: Unveiling of the \'Vikram\' Large Language Models

In a major step toward technological self-reliance, Bengaluru-based startup Sarvam AI unveiled two indigenous Large Language Models (LLMs) named Vikram at the AI Impact Summit 2026. Named in honor of the father of the Indian space program, Dr. Vikram Sarabhai, these models represent India\'s ambition to create a \'sovereign AI\' ecosystem that caters specifically to its unique linguistic and cultural landscape. • Dual-Model Strategy: Sarvam AI introduced a 35-billion parameter model optimized for real-time conversational tasks and a flagship 105-billion parameter model designed for advanced reasoning and complex agentic workloads.• Linguistic Mastery: Unlike global models predominantly trained on English, Vikram is built from scratch to support all 22 official Indian languages, featuring voice-first optimization to bridge the digital divide for non-English speaking populations. • Public-Private Partnership: The development was supported by the IndiaAI Mission, providing the startup with subsidized access to high-performance computing (GPUs) and aligning with the government\'s goal of reducing dependency on foreign AI platforms. • Efficiency & Scaling: The models utilize a Mixtureof-Experts (MoE) architecture, which activates only a fraction of total parameters (e.g., 1 billion for the 35B model) during inference, significantly lowering the cost and energy required for population-scale deployment. • Open-Source Commitment: To foster a domestic innovation hub, Sarvam AI has committed to releasing these models as open-source, allowing Indian developers, researchers, and government agencies to build localized applications on top of the Vikram foundation. • Sovereign Data Security: By developing and deploying these models within India, the initiative ensures that sensitive citizen data remains under domestic jurisdiction, addressing critical concerns regarding data sovereignty and national security. Definitions & Technical Concepts • Large Language Model (LLM): An AI system trained on vast datasets to understand, generate, and manipulate human language. It uses deep learning to predict the next \'token\' in a sequence. • Parameters: Often described as the \'brain cells\' of an AI, parameters are the internal variables the model learns during training. Generally, more parameters allow for more complex reasoning. • Mixture-of-Experts (MoE): A neural network architecture where only specific \'expert\' parts of the model are activated for a given task, making large models much faster and cheaper to run. • Token Fertility: A measure of how many \'tokens\' (units of text) an AI needs to process a single word. Vikram is designed for high efficiency in Indic scripts, which typically have high fertility in Western models like GPT. Constitutional & Legal Context • Digital Public Infrastructure (DPI): Sovereign AI is increasingly viewed as a new layer of India\'s DPI, alongside UPI and Aadhaar, intended to democratize access to high-end technology. • IndiaAI Mission: A flagship initiative under the Ministry of Electronics and Information Technology (MeitY) with an outlay of over ₹10,300 crore to build a comprehensive AI ecosystem including compute capacity, datasets, and indigenous models. • Data Sovereignty: While not explicitly a constitutional term, it relates to the right of a nation to govern data generated within its borders, linked to the Right to Privacy (Article 21) as per the Puttaswamy judgment. Important Keypoints for Analysis • Overcoming Data Scarcity: A major challenge for Indic AI is the lack of high-quality digital text in regional languages. Sarvam AI addressed this by curating a high-quality corpus of 2 trillion tokens across 10 major languages. • Global Benchmarking: Despite being smaller in size than global giants like DeepSeek (600B) or GPT4, Vikram models have shown competitive performance in mathematical reasoning and OCR (Optical Character Recognition) for Indian scripts.• Voice-First Economy: In a country with varying literacy levels, voice-enabled AI in local dialects (like the Bulbul text-to-speech model) is seen as the primary vehicle for delivering government services and education. Conclusion The launch of the Vikram models marks a pivot from being a consumer of AI to a creator of foundational technology. While the commercial sustainability of open-source models remains a challenge, the strategic value of having a \'Made in India\' LLM lies in its ability to serve 1.4 billion people in their own languages while maintaining control over the underlying intelligence. UPSC Relevance • General Studies II: Government policies and interventions for development in various sectors; Governance and E-applications. • General Studies III: Science and Technology—developments and their applications and effects in everyday life; Indigenization of technology and developing new technology; IT and Computers. • Economic Development: Role of AI in achieving a $5 trillion economy and the significance of the IndiaAI Mission in promoting startups.