Sarvam AI LLMs, IndiaAI Mission and the Rise of Indigenous Large Language Models

At the AI Impact Summit 2026, the Bengaluru-based startup Sarvam AI released two Large Language Models (LLMs).

• The two models were trained on 35 billion and 105 billion parameters, respectively, and were less power- and compute-intensive than comparable models.

About Large Language Models (LLMs)

• Large language models (LLMs) are advanced AI systems designed to understand and generate human-like text. 
  • They learn from vast amounts of written data to predict what comes next in a sentence or to create coherent responses to questions. 
• Architecture and Training: LLMs use deep learning with transformer architectures, like Generative Pre-trained Transformer (GPT), designed for processing sequential text data. 
• • They feature multiple neural network layers and an attention mechanism for context understanding.

Training of Large Language Models (LLMs)

• Training Process: LLMs are trained on massive clusters of Graphics Processing Units (GPUs), which provide the computational power required to process vast amounts of data.
  • The model learns to predict the next word in a sentence based on the context provided by previous words.
  • Tokenization and Embeddings: Words are broken down into tokens, which are then converted into numerical embeddings representing the context.
  • Massive Text Corpora: LLMs are trained on extensive text data, allowing them to learn grammar, semantics, and conceptual relationships.
  • Learning Techniques: They use zero-shot and self-supervised learning to generalise from the data.
    • Zero-shot learning refers to a model’s ability to handle tasks or make predictions about data it has not seen during training.
  • Enhancing Accuracy: Performance is improved through prompt engineering, fine-tuning, and reinforcement learning with human feedback (RLHF) to address biases and inaccuracies.

Explore Combo (GS+Optional) Courses

Explore Combo (GS+Optional) Courses

Challenges in Training LLMs

• Limited Capital: Since capital is scarce, efforts to train an LLM by Indian firms targeting Indian users can be challenging, especially if there is no immediate business use case for doing so.
  • For example, training a 70-billion-parameter LLM can cost around $6 million, a prohibitive amount for early-stage Indian startups without assured near-term returns.
• High Capital Intensity: Training and operating LLMs requires expensive GPU clusters and massive electricity consumption, running into millions of dollars.
  • For Example: Training GPT-3 cost over $4–5 million in compute, while GPT-4 reportedly required tens of millions of dollars and thousands of GPUs running for months.
• Scarcity of Indian Language Data: Internet data is dominated by English, European, Korean, and Japanese content, leaving Indian languages underrepresented.
  • For Example, English makes up over 50% of web content, while most Indian languages each account for less than 1%, leading to minimal representation in datasets like Common Crawl.
• Performance Gap in Indian Languages:  Due to limited native datasets, LLMs often perform poorly in Indian languages compared to English.
• Higher Token Consumption:  Many models translate Indian language inputs into English for better processing and then translate outputs back, increasing token usage and inference costs.
  • For Example: A 10-word English sentence may use around 12–15 tokens, whereas the same sentence in Hindi (Devanagari script) can consume 20–25 tokens due to tokenisation inefficiencies. 

About Mixture of Experts (MoE) 

• Mixture of Experts (MoE) is a way of designing AI models so that only the necessary parts of the model are used for each question, instead of using the whole model every time.
• For Example: 
  • Imagine a school with many teachers (experts).
  • If a student asks a maths question, only the maths teacher answers, not the history or science teachers.
• Similarly:
  • In a normal AI model, all parts work for every question, which uses a lot of power and money.
  • In an MoE model, only a few specialised parts are activated, making it faster and cheaper.

Way Forward

• Expand Indian Language Datasets: Create large, high-quality, annotated corpora in Hindi, Tamil, Bengali, Marathi and other Indian languages through public–private partnerships and initiatives like Bhashini.
• Focused Sectoral Models: Develop smaller, domain-specific LLMs for governance, education, healthcare, agriculture, and law instead of only competing with frontier global models.
• Industry Academia Collaboration: Strengthen partnerships between IITs, IIITs, startups, and MeitY to build skilled AI talent and research depth.
• Energy Efficient Architectures:  Adopt approaches like Mixture of Experts (MoE) and model compression to reduce training and inference costs

Check Out UPSC CSE Books

Visit PW Store

DOWNLOAD PDF