India Combines Open AI Models and Local Expertise to Drive Innovation

India’s ambitions in artificial intelligence (AI) were prominently showcased at Google’s annual I/O Connect event held in Bengaluru in July 2023. With more than 1,800 developers in attendance, discussions revolved around enhancing AI capabilities tailored to the country’s rich linguistic tapestry. India, home to 22 official languages and numerous dialects, faces a significant challenge in creating AI systems that can effectively operate across this multilingual landscape.

During the event, various startups highlighted innovative solutions to this challenge. Notably, Sarvam AI presented its multilingual model, Sarvam-Translate, built on Google’s open-source large language model (LLM), Gemma. Similarly, CoRover showcased BharatGPT, a chatbot designed to streamline public services for organizations like the Indian Railway Catering and Tourism Corporation (IRCTC). Google announced that startups including Sarvam, Soket AI, and Gnani are collaborating to refine the next generation of AI models using Gemma, while also participating in the government-led 10,300 crore IndiaAI Mission aimed at developing indigenous foundational models.

At first glance, this dual approach may appear contradictory. The initiative emphasizes creating AI models from the ground up, specifically trained on Indian data. Yet, the reality is that building competitive models in isolation is resource-intensive. Given the limited availability of high-quality training datasets and the urgent market demand, these startups are opting for a pragmatic method. They are fine-tuning existing open-source models to address immediate needs while simultaneously developing the infrastructure and expertise required for more independent models in the future.

Fine-tuning involves adapting an already established large language model, trained on extensive general data, to excel in specific local contexts. Project EKA, an open-source initiative led by Soket, exemplifies this strategy. In collaboration with IIT Gandhinagar, IIT Roorkee, and IISc Bangalore, EKA aims to create a sovereign LLM by sourcing training data and infrastructure locally. A 7 billion-parameter model is expected within four to five months, with plans for a 120 billion-parameter model developed over a ten-month cycle.

Abhishek Upperwal, co-founder of Soket AI, outlined their strategy: “We’ve mapped four key domains: agriculture, law, education, and defence. Each has a clear dataset strategy.” A critical feature of the EKA initiative is that it operates independently of foreign infrastructure, with training conducted on India’s GPU cloud. The resulting models will be open-sourced for public access.

CoRover’s BharatGPT employs a similar dual strategy, using a fine-tuned model to provide conversational AI services in multiple Indian languages to government clients. Founder Ankush Sabharwal emphasized the necessity of a base model that could be quickly adapted for various applications. “We needed a model that could respond swiftly to local needs, while simultaneously building our foundational capabilities,” he stated.

Amlan Mohanty, a technology policy expert, described India’s approach as a strategic experiment. It aims to leverage models like Gemma for rapid deployment while pursuing long-term autonomy. “It’s about reducing dependency on adversarial countries and ensuring cultural representation,” he noted. Despite reaching out to Sarvam and Gnani for insights on their use of Gemma, neither responded.

For India, the development of AI capabilities transcends national pride; it addresses pressing local challenges. Consider a migrant from Bihar visiting a rural clinic in Maharashtra. If the doctor speaks Marathi and the AI tool explains medical findings in English, vital nuances may be lost. A health worker in Bihar requires an AI tool that understands local medical terms in Maithili, just as a farmer in Maharashtra needs advice tailored to local irrigation practices.

These scenarios illustrate the critical need for AI systems that accurately reflect local languages, contexts, and values. Fine-tuning existing models offers Indian developers a means to address immediate requirements while constructing the datasets necessary for a truly sovereign AI infrastructure. This layered strategy may represent one of the most effective pathways forward, utilizing open tools to progressively build local capabilities.

The IndiaAI Mission serves as a national response to the growing geopolitical landscape surrounding AI. As these systems become integral to sectors like education, agriculture, and governance, reliance on foreign platforms poses risks concerning data security and control. The abrupt suspension of Microsoft’s cloud services to Nayara Energy due to European Union sanctions on its Russian-linked operations underscored this concern.

In addition to reducing reliance on foreign technology, the development of sovereign AI is vital for ensuring that India’s unique values and regulatory frameworks are accurately represented. Most global AI models are primarily trained on English-speaking and Western datasets, rendering them inadequate for the diverse linguistic and contextual realities of India.

Mohanty argues that sovereignty in AI is not about isolation but rather about control over the infrastructure and decision-making. “Sovereignty is fundamentally about choice and dependencies. The more options you have, the greater your sovereignty,” he explained.

Despite the progress made, the lack of high-quality training data, particularly in Indian languages, remains a significant obstacle. Manish Gupta, director of engineering at Google DeepMind India, highlighted that an internal assessment revealed that 72 of India’s spoken languages, each with over 100,000 speakers, had virtually no digital presence. To address this issue, Google is collaborating with the Indian Institute of Science to collect voice samples across hundreds of districts, with the first phase capturing over 14,000 hours of speech data in 59 languages.

As India continues to develop its AI ecosystem, the dual-track strategy of leveraging open models while cultivating local expertise could provide a roadmap for other nations facing similar challenges. Countries in the Global South, such as Singapore, Vietnam, and Thailand, are already exploring comparable methods. By 2026, when India’s sovereign LLMs are expected to be operational, the hope is that the dual tracks will converge, leading to a robust, self-sufficient AI landscape.

As India and other nations navigate the complexities of AI development, the emphasis will remain on converting borrowed support into a complete, sovereign infrastructure before external conditions change. The journey towards AI independence is not merely a sprint but a marathon, requiring both immediate action and long-term vision.