According to BCG India’s AI revolution reached new heights in 2024 with innovative models like Sarvam 2B, BharatGen, and Devika, combining modern technology with cultural richness. How is India’s AI ecosystem shaping the future of technology for the world?

India is leading in AI adoption at 30%, surpassing the global average of 26%, according to a report by Boston Consulting Group. India is leading in Artificial Intelligence (AI) adoption, according to new research by Boston Consulting Group (BCG), which pegged that 30 per cent of Indian companies are maximising value through the use of such emerging technology

AI's role in India's economy could be key to achieving the goal of building a $1 trillion digital economy by 2028. According to Google's report, AI Opportunity Agenda for India, the strategic use of AI could generate at least ?33.8 lakh crore (USD 407.23 billion) in economic value by 2030. This growth would come from higher productivity, technological advancements, and more efficient operations across industries like agriculture, healthcare, and finance.

Agriculture, one of India’s most crucial industries, can greatly benefit from AI-powered tools. For instance, AI-driven mobile platforms like Agrostar use cloud technologies to provide small farmers with multilingual advice, improving crop yields and promoting sustainable farming practices. In healthcare, AI’s ability to predict trends can help prevent diseases and optimise medical resource allocation, effectively addressing public health challenges.

Similarly, Divinet is a start-up focused on creating multi-play technology for remote learning and teaching. It incorporates AI into remote learning platforms to personalise and scale educational solutions, opening new opportunities for learners and educators alike.

Dr. Vijay Bhatkar, one of the co-founders of Divinet and the architect of India’s national supercomputing initiative, has led the creation of the Param 8000 and Param 10000 supercomputers. These systems provided the computational power necessary for complex AI applications, laying a strong foundation for AI advancements in India.

Dr. Bhatkar has also contributed to India’s AI space with his pioneering work in broadband technology, including the introduction of Wi-Fi and WiMAX in India. 

In other words, India is innovating in developing the digital infrastructure needed for AI research and its practical applications.

AI-Driven Governance: Key Indian Initiatives

India has been proactively incorporating Artificial Intelligence (AI) across various public services and infrastructures through several significant initiatives:

  1. Applied AI research centre (INAI) in Telangana: In partnership with Intel, IIIT Hyderabad, and the Public Health Foundation of India, this Hyderabad-based centre leveraged AI to address healthcare and smart mobility challenges. It aims to enhance the accessibility of public healthcare and facilitate smart mobility, particularly in rural areas and difficult terrains.
  2. US-India AI initiative: Launched by the Indo-US Science and Technology Forum (IUSSTF) on March 18th, this initiative fosters AI growth through idea exchange, research and development exploration, and bilateral collaboration. The focus areas include energy, health, and agriculture, aiming to harness diverse data inputs to enhance national development.
  3. MCA 3.0 portal: The Ministry of Corporate Affairs has upgraded its portal to simplify regulatory filings and improve compliance monitoring with AI and advanced analytics. This version aims to enhance the ease of doing business and establish a well-organised business regulatory environment.
  4. Responsible AI for youth: Managed by the National e-Governance Division of MeitY, this program targets government school students, aiming to bridge the AI skills gap. It focuses on practical skills development over theoretical knowledge to boost employability and encourage students to apply AI in addressing significant social challenges.
  5. AI portal: A joint initiative by MeitY and NASSCOM, this portal serves as a central hub for AI-related resources, updates, and community engagement in India. It supports AI enthusiasts, experts, and professionals by providing access to the latest news, expert articles, and opinion pieces.
  6. National Research Foundation: Established under the National Education Policy (NEP), the NRF aims to synergise R&D, industry, and academia. It enhances the research lifecycle, accelerates application development, and ensures the practical longevity of AI solutions.

Additionally, under the aegis of the Digital India Corporation, the ‘IndiaAI’ Independent Business Division is executing several components to boost AI innovation and application:

  • IndiaAI Compute Capacity: Establishing a scalable AI compute infrastructure through public-private partnerships, aiming for a network of over 10,000 GPUs and creating an AI marketplace for resources.
  • IndiaAI Innovation Centre: Developing and deploying indigenous Large Multimodal Models (LMMs) and domain-specific foundational models in critical sectors.
  • IndiaAI Datasets Platform: Facilitating access to quality non-personal datasets for AI innovation through a unified data platform.
  • IndiaAI Application Development Initiative: Promoting AI applications in critical sectors, focusing on developing and scaling impactful AI solutions.
  • IndiaAI FutureSkills: Expanding AI education at various levels and setting up Data and AI Labs in Tier 2 and Tier 3 cities.
  • IndiaAI Startup Financing: Supporting deep-tech AI startups with streamlined access to funding for AI projects.
  • Safe & Trusted AI: Implementing Responsible AI projects, including the development of indigenous tools, self-assessment checklists, and governance frameworks to ensure the ethical use of AI.

India’s first open-source foundational model: Sarvam 2B

Bengaluru-based startup Sarvam AI has launched Sarvam 2B, India’s first open-source foundational model created entirely from the ground up. This achievement shows the startup’s dedication to developing independent AI models that address specific national needs and unique use cases. Supported by a $41 million investment from prominent firms such as Lightspeed, Peak XV Partners, and Khosla Ventures, Sarvam AI is establishing itself as a leader in India’s AI innovation.

Sarvam 2B was trained on a massive dataset of 4 trillion tokens, enabling it to understand and process instructions in 10 Indian languages: Hindi, Tamil, Telugu, Malayalam, Punjabi, Odia, Gujarati, Marathi, Kannada, and Bengali. This capability makes it stand out from other language models and is ideal for tasks such as translation, summarisation, and understanding colloquial expressions in Indian languages.

Vivek Raghavan, co-founder of Sarvam AI, stated that Sarvam 2B belongs to a category of Small Language Models (SLMs), similar to Microsoft’s Phi series, Llama 3 (8 billion), and Google’s Gemma models. He explained, “This is the first open-source foundational model trained on an internal dataset of 4 trillion tokens by an Indian company, with computation in India, and with efficient representation for 10 Indian languages.”

The model is available on Hugging Face, encouraging further research and enabling developers to create new applications. Unlike other foundational models in India, such as Tech Mahindra’s Project Indus and Krutrim, Sarvam 2B stands out for being fully open-source, demonstrating the startup’s focus on transparency and accessibility.

Sarvam AI’s innovations are not limited to text-based language models. In August 2024, the startup introduced Shuka 1.0, India’s first open-source audio language model, at an event held at ITC Gardenia in Bengaluru. Shuka 1.0 enhances the capabilities of the Llama 8B model by supporting voice input in Indian languages and providing text output. It surpasses the accuracy of leading models like Whisper combined with Llama 3 and processes audio using audio tokens for seamless language model integration.

Sarvam AI has prioritised voice-enabled generative AI. As co-founder Raghavan mentioned, “We believe consumers in India will prefer using generative AI through voice mode rather than text.” The model is said to be six times faster than Whisper + Llama 3 and offers higher accuracy across 10 Indian languages.

Sarvam Agents: Voice-Based AI for Business Solutions

Sarvam AI has also launched Sarvam Agents, multilingual, voice-based AI tools that go beyond traditional chatbots. These agents are designed to address specific business needs and can be accessed through platforms like telephony, WhatsApp, and mobile apps. With an affordable pricing model starting at INR 1 per minute, these agents are particularly beneficial for contact centres and sales teams aiming to adopt advanced AI solutions.

Raghavan highlighted the contextual capabilities of these agents, which make them unique compared to existing conversational AI systems. He explained, “Sarvam Agents can respond contextually. For example, if users request information from a specific page, the agent can reply with that context in mind, unlike standard call-centre systems that begin the interaction from scratch.”

Additional Offerings: Sarvam Model APIs

To enhance its open-source models, Sarvam AI is offering APIs for its proprietary Indic models, which play a key role in powering Sarvam Agents. These include:

  • Speech-to-Text Model: Converts spoken Indian languages into English with exceptional accuracy, outperforming traditional Automatic Speech Recognition (ASR) systems.
  • Text-to-Speech Model: Transforms written text into speech, offering various voice options in multiple languages to suit different preferences.
  • Parsing Model: Specialised in extracting data from complex documents like financial statements, ensuring high precision in document analysis.

Other best Indian AI models launched in 2024

BharatGen: A Homegrown Initiative with Cultural Depth

The Indian government took a major step towards promoting local innovation by launching BharatGen, the country’s first state-funded project to create multimodal large language models (LLMs). Its tagline, “GenAI for Bharat, by Bharat,” reflects its focus on designing AI that understands and respects India’s diverse culture and languages.

One standout model from BharatGen is e-vikrAI, which uses Vision Language Models (VLMs) to improve the e-commerce experience. With this model, sellers can upload product images and receive detailed titles, descriptions, and features without manual input. Its cultural awareness and fluency in Indic languages make it an invaluable tool for small businesses, particularly those in regional markets.

Nemotron-4-Mini-Hindi-4B: NVIDIA’s Regional Focus

During his visit to India, NVIDIA CEO Jensen Huang unveiled the Nemotron-4-Mini-Hindi-4B model, a compact yet powerful Hindi language model tailored to meet regional business needs.

Part of NVIDIA’s NIM microservice, this model is designed for deployment on NVIDIA GPU-accelerated systems, optimising performance across a range of applications.

Tech Mahindra became the first company to adopt the model, launching Indus 2.0, a platform focusing on Hindi and its dialects. With 4 billion parameters, Nemotron-4-Mini-Hindi-4B is derived from its larger sibling, the 15-billion-parameter Nemotron-4. It was trained using a combination of real-world Hindi data and synthetic datasets, including English.

Fine-tuned with NVIDIA NeMo, the model achieves leading scores on accuracy benchmarks for AI models with up to 8 billion parameters. Packaged as a microservice, it supports various industries, including education and healthcare.

Krutrim by Ola: A Multilingual Breakthrough

Ola Krutrim, the AI division of Ola led by Bhavish Aggarwal, has made remarkable progress in advancing Indian AI. At the Sankalp 2024 event, Aggarwal highlighted the achievements of Krutrim over six months, its growing influence on Indian developers, and its ambitious plans for the future.

Since its launch, Krutrim Cloud has gained significant traction, processing approximately one trillion tokens, with 250 billion API calls and 25,000 developers actively using the platform.

Krutrim is now focused on building India’s first AI cloud infrastructure, fully hosted in its own data centres and powered by its proprietary Bodhi chips. The first version of these AI chips is expected by 2026, with an upgraded Bodhi 2 planned for 2028. To celebrate its success, Krutrim Cloud will remain free for users until Diwali this year, encouraging more developers to explore its capabilities. Aggarwal announced over 50 new services for the Krutrim Cloud platform, designed to cater specifically to Indian developers. These include:

  • AI Pods
  • AI Studio
  • A comprehensive language hub supporting multimodal translation and customer experience AI.

Since its launch, Krutrim-7B-chat, a large language model (LLM) trained in ten Indic languages, has been available on the Databricks Marketplace for free. According to Gautam Bhargava, VP and Head of AI Engineering at Krutrim, the platform now features the most effective Indic tokeniser compared to competitors like GPT-4o and Gemini, although it slightly trails Llama 3.1.

Edtech platform Unacademy has already adopted Krutrim’s translation API to enhance its language learning programmes. An upcoming image recognition feature was also showcased, further expanding the functionality of Krutrim’s AI solutions.

Surya OCR: Transforming Document Processing

Surya OCR is an advanced optical character recognition (OCR) toolkit created by Vik Paruchuri. The recently launched version 2 delivers enhanced accuracy and outperforms tools such as Tesseract and Google Cloud OCR.

Surya OCR is capable of recognising diverse document layouts, including tables, images, and headers, while maintaining the original structure during processing.

It is easy to install via pip and requires Python 3.9+ along with PyTorch. Model weights are downloaded automatically during the first run, simplifying setup. A reliable API is also available for those who prefer hosted solutions, supporting various document formats for consistent OCR performance.

The toolkit includes a robust command-line interface for tasks such as OCR, text detection, layout analysis, and reading order detection. Surya OCR can also be integrated into Python scripts for custom workflows, enabling advanced preprocessing and OCR pipelines.

Everest 1.0 by Hanooman: Multilingual Mastery

On October 24, Hanooman AI launched Everest 1.0, a multi-lingual and multi-modal foundational large language model (LLM), at the NVIDIA AI Summit in Mumbai. This milestone underscores Hanooman AI’s commitment to advancing India’s position as a leader in AI technology. Everest 1.0 currently supports 35 languages, with plans to extend this capability to 90 languages in the coming months. This comprehensive language support aims to facilitate diverse applications across linguistic and cultural boundaries, bolstering the model’s utility for a global audience.

Everest 1.0 has been designed to perform a wide array of tasks, including text generation, image creation, code writing, voice generation, and document analysis. The platform is deployed on Yotta infrastructure and utilises cutting-edge NVIDIA GPUs for training and operational efficiency. Hanooman AI describes Everest 1.0 as a significant milestone that contributes to India’s vision of becoming a “global AI factory,” showcasing the country’s technological prowess in creating complex, multi-functional AI systems.

Chitralekha: AI-Powered Video Transcreation

AI4Bharat, in collaboration with EkStep, has introduced Chitralekha, an open-source AI-powered platform designed to simplify video transcreation. This platform manages the entire process of converting videos from one language to another, including transcription, translation, and voice-over generation. Chitralekha supports English and a variety of Indian languages, making it ideal for creators working with diverse linguistic audiences.

Chitralekha includes a workforce management system that streamlines tasks from transcription to voice-over. This structured approach helps users efficiently manage the video transcreation process. Additionally, users can edit mp3 audio transcriptions, expanding the platform’s utility beyond videos.

Extensive Language Support:

  • Transcription: Available in 12 languages.
  • Translation: Covers all 22 official Indian languages.
  • Voice-over: Automated generation in 14 languages, perfect for creating multilingual educational and informational content.

Open-Source Tools:
The platform utilises Subplayer, an open-source subtitling tool available at subplayer.js.org, for creating subtitles. Chitralekha also incorporates models developed by AI4Bharat, including:

  • IndicASR for transcription
  • IndicTrans for translation
  • IndicTTS for voice-over generation

While its current voice-over feature supports single-speaker videos, future updates aim to include multi-speaker content.

Chitralekha prioritises accuracy and customisation. It offers a detailed verification and correction process for machine-generated transcripts and subtitles, allowing users to manually refine them for top-quality results. Subtitles can be downloaded in various formats, such as srt, vtt, txt, and docx, while interactive word-level aligned subtitles are available in ytt format, highlighting each word as it is spoken. 

Chitralekha’s intuitive interface offers features like:

  • Adjustable font sizes
  • Video playback speed controls
  • Theme selection
  • Customisable subtitle positioning

Airavata: Hindi Language Excellence

Airavata is an open-source, instruction-tuned LLM for Hindi, developed by AI4 Bharat at IIT Madras. Released in 2024, it was created by fine-tuning Sarvam AI’s OpenHathi model using diverse instruction-tuning datasets in Hindi.

Named after the Sanskrit word for “elephant,” Airavata excels at handling a wide range of tasks in Hindi. It was trained with human-curated, licence-friendly datasets, including translations of English instruction datasets, ensuring sustainability and avoiding licensing restrictions.

Sutra: Efficiency in Multilingual AI

Two AI, a startup focused on the untapped markets of southern Asia, has introduced SUTRA, a multilingual AI model designed to compete with top-tier models like GPT-4. Developed to support more than 30 languages, SUTRA is especially notable for its proficiency in South Asian languages such as Gujarati, Marathi, Tamil, and Telugu. Accompanying this model is ChatSUTRA, a free web-based chatbot that operates using SUTRA’s capabilities.

SUTRA's architecture features two mixture-of-experts transformers: a concept model and an encoder-decoder for translation. The concept model, trained to predict subsequent tokens, utilises publicly available datasets in languages with abundant data, including English. In parallel, the translation model was trained on 100 million human- and machine-translated conversations spanning multiple languages, allowing it to map concepts to similar embeddings across the linguistic spectrum.

The combined training approach involves the translation model’s encoder feeding into the concept model, which then passes its output to the decoder of the translation model. During inference, this system processes input text through the encoder, generates an initial embedding, refines it via the concept model, and outputs translated text through the decoder. This intricate design supports efficient cross-linguistic understanding and translation.

SUTRA is accessible via an API in three versions:

  • SUTRA-Pro: High performance, priced at $1 per 1 million tokens.
  • SUTRA-Online: Internet-connected, also at $1 per 1 million tokens.
  • SUTRA-Light: Lower latency, available at $0.75 per 1 million tokens.

Devika: Autonomous AI Engineering

Devika, an innovative AI software engineer, was created by Mufeed, a young developer from Kerala. Inspired by Cognition Labs in the US, which developed the world’s first fully autonomous AI software engineer, Devin, Mufeed aimed to build an AI that would address Indian and global needs while showcasing an Indian identity. 

Devika’s capabilities are powered by state-of-the-art AI models, including Claude 3, GPT-4, GPT-3.5, and locally trained large language models (LLMs). This robust foundation allows Devika to handle complex software tasks with efficiency and precision.

Devika is designed to handle a range of tasks in software development, including:

  • Understanding and executing human instructions: The AI is capable of comprehending complex, high-level human instructions and converting them into actionable coding steps.
  • Code generation and debugging: Devika autonomously generates software code and performs bug-fixing without requiring user intervention.
  • Research and execution: It can conduct research related to assigned tasks and complete them efficiently, mimicking the behaviour of an experienced software engineer.
  • Feedback loop with agentic models: Devika incorporates 12 agentic models that function within a feedback loop to optimise its performance and reliability.

The Impact of Indian AI Models in 2024

The AI innovations of 2024 highlight India’s commitment to culturally and linguistically aware technology. These models have transformed sectors such as healthcare, education, and e-commerce, empowering businesses and enhancing customer experiences.

From BharatGen’s e-vikrAI to the resource-efficient Sarvam-1, each model tells a story of progress. India’s strategic investments, academic collaborations, and entrepreneurial spirit are driving its AI ecosystem forward, proving that it is not only participating in but also leading global AI advancements. These models celebrate India’s diversity while setting new benchmarks for inclusive and impactful technology.