Indigenous Large Language Model Vikram Developed by SarvamAI

Indigenous Large Language Model Vikram Developed by SarvamAI | India AI Innovation

SarvamAI has developed Indigenous Large Language Model Vikram designed to strengthen India’s AI ecosystem. Explore its features, objectives, technological significance, and impact on India’s digital sovereignty and AI innovation landscape.

Table of Contents

Context

Bengaluru-based Sarvam AI unveiled two language models named Vikram at the AI Impact Summit.

About the Vikram

It is a s a sovereign AI platform designed to provide specialized, intelligent solutions tailored for the Indian context,
Open source: the models will be open source.
- Open-source LLMs have source code that is open to the public, making it available to anyone who wants to access it.
It aims to reduce reliance on foreign AI systems by focusing on regional language capabilities and local data
Purpose: : The platform aims to bridge the language gap in AI, offering solutions that understand India’s diverse cultural and linguistic landscape
It is available in two versions — 35-billion parameters and 105-billion parameters.
- 35B model →Suitable for scalable deployment, faster inference, lower compute cost.
- 105B model →Higher reasoning capability, better contextual depth, advanced generative performance.
With tens of billions of parameters, it falls in the category of advanced large language models (LLMs)

About Large Language Models (LLMs)

● They are advanced AI systems designed to understand, process, and generate human language.
They are called “large” because they contain millions or billions of parameters trained on massive text data.

● Examples include GPT models (used in ChatGPT), BERT, PaLM etc

Core Technology – Transformers

● LLMs are powered by transformer neural networks, introduced by Google in the 2017 paper “Attention Is All You Need.”

How LLMs Work

● Embeddings: Words are converted into numerical vectors (embeddings) that capture meaning and relationships.

● Attention Mechanism: Helps the model understand context and relationships between words in a sentence.

Training Process

Pre-training

● Trained on massive internet text data.

● Uses unsupervised learning.

● Learns language patterns.

● Very expensive and resource-intensive.

Fine-tuning

● Model is adapted for specific tasks.

● Often uses Reinforcement Learning from Human Feedback (RLHF).

● Makes LLMs useful for real-world applications.

Read More Notes
Environment Notes	Art and Culture Notes
Science and Tech	History Notes
Geography Notes	Indian Polity Notes
General Knowledge	International Relation