When Sam Altman visited India last year, he said it would be impossible for a startup to compete with OpenAI at training foundation models with $10 million in the bank. The comment made major headlines, with CP Gurnani, the former CEO of Indian IT firm Tech Mahindra, ambitiously saying that the challenge to build generative AI natively in India was accepted.
Fast forward to early 2024, India, which is known for its technology talent and companies, is well on its way with generative AI. However, the interesting part is the first Indian player making a concrete move to take on OpenAI’s GPT models is not Tech Mahindra but — you guessed it — a startup founded by Bhavish Aggarwal, who also founded ride-hailing company Ola Cabs to take on Uber.
Ola Krutrim – which means “artificial” – debuted its first language model, Krutrim base, and a chatbot built on top of it last month while detailing plans to take it mainstream very soon. Other players, including Tech Mahindra and Reliance Industries, are also in the race, trying to catch up.
The race to deliver localized experiences
While foundation models such as OpenAI’s GPT family and Meta’s Llama do a pretty good job at generating language, answers and code, they can sometimes struggle to handle queries in non-English languages, particularly low-resource ones (with a smaller digital footprint).
To address this and power more localized experiences, technology companies in different countries, including South Korea, Finland, and China, have started training proprietary models with an approach of increasing the representation of local languages and cultural contexts in their training data.
The same challenge also impedes India’s generative AI ambitions. However, the problem is multifold bigger in this case. The country is home to 1.4 billion people, or nearly 18% of the world’s population, and has 22 officially recognized languages, 1,600+ dialects and 19,200 unofficial dialects. Training a model to encompass all of it is a task in itself – and certainly a capital-intensive one (as Altman suggested).
After offering ride-hailing services and selling electric vehicles, Aggarwal incorporated Krutrim in April 2023 to take on this challenge. The company raised $24 million in debt from Matrix Partners and trained Krutrim based on two trillion tokens. This, the entrepreneur touted at launch, includes the largest representation of Indic languages, 20 times more than any other model.
“Krutrim has Indian ethos, natively. It generates text and code with an innate sense of Indian cultural sensibilities and relevance,” he said.
In its current form, Ola’s model understands 20 Indian languages and generates 10, including Hindi and English.
According to the company, its performance across Indic languages is already better than GPT-4 but English quality performance remains behind (it is expected to improve in the coming months.)
The startup is moving in phases and has multiple developments in the pipeline, including support for all officially recognized Indic languages and a Pro version of the model for complex problem-solving with support for text, vision and speech.
In addition to the models, which would be provided to businesses, Aggarwal and the team have built a ChatGPT-like chatbot experience for the Indian audience. However, it is not open to the public at this stage. The company is also doing R&D on the hardware front to build its AI supercomputer.
Big guns playing catchup
While it remains to be seen how Krutrim’s models pan out in the real world, when developers and consumers begin to use them, the company has positioned itself as one of the first Indian players to cover all the bases in the much-hyped generative AI space.
The other notable companies that are playing catch up are Tech Mahindra and billionaire Mukesh Ambani’s Reliance Industries.
Tech Mahindra, under CP Gurnani’s leadership, started working on an open-source large language model (LLM) under The Indus Project in August 2023 and recently launched it for internal beta testing.
This offering is slated to debut in February 2024 and is said to be a pure Hindi LLM with 539 million parameters and 10 billion Hindi + dialect tokens. Even in this case, not all languages are supported.
“In the first phase, we will be creating the LLM for Hindi language and 37+ dialects, and then move ahead in a phased manner to cover other languages and dialects,” the company noted on its website.
On the other hand, Reliance Industries, which led the 4G wave in India with Jio and has backers like Google, Meta and Intel, appears to be moving a tad slower in the race for AI.
The company announced a plan to build language models for India at its AGM last year. It subsequently partnered with Nvidia to gain access to the GH200 superchip and build AI infrastructure more powerful than the fastest supercomputer in India. Now, it is working with a team at the Indian Institute of Technology-Bombay to bring the project, dubbed Bharat GPT, to life.
While not many details have been shared, it appears that Reliance plans to bring the GPT offering across its customer-facing products and services, including those offered by Jio. It’s unclear if the company will launch a separate, ChatGPT-like consumer-facing chatbot or not.
Along with Reliance and TechM, Bengaluru-based Sarvam AI, which recently came out of stealth with $41 million in funding, has also garnered significant attention.
The startup has built a 7 billion parameter Indic language model, based on Llama2, and plans to launch an enterprise-centric platform to help companies build generative AI apps using it.
Google-backed Corover also claims to have built an Indic language model supporting 22 languages for its platform for conversational enterprise chatbots.
Better experiences with generative AI
As the ecosystem evolves, more players emerge and technology matures, more sophisticated closed and open-source Indic language models are expected to take shape in the country. All this will not only improve internal enterprise workflows but also lead to better applications for organizations operating across different sectors.
For instance, Tech Mahindra notes Indus Project’s LLM can lead to the development of a digital helper for more than 140 million farmers, providing them with the required information on loans, pesticides and other agriculture-related aspects in their preferred language.
It could also power healthcare and finance kiosks to decipher speech in local dialects and provide useful information in a matter of seconds. The possibilities are endless.
Beyond this, it will also be interesting to see how these models fare against their global counterparts in terms of performance, including market leaders like OpenAI, which is closing towards GPT-4.5, and Google, which recently debuted the Gemini series of models.
TechForgePulse's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.