Ola co-founder and CEO Bhavish Aggarwal rolled out Krutrim, a ‘made in India’ large language model and generative AI platform on the lines of OpenAI’s ChatGPT and Google’s Bard. The platform is developed and owned by Krutrim Si Design, the AI project of Aggarwal. It is trained on 2 trillion tokens or pieces of textual information, having the most significant representation of Indian data. 

Krutrim means artificial in Sanskrit and comes in a base model and a pro model. Like ChatGPT, this primary version can respond to prompts or queries from the public. It can comprehend around 22 Indian languages and generate text in 10 languages like Marathi, Hindi, Bengali, Tamil, Kannada, Telugu, Odia, Gujarati and Malayalam. Krutrim Pro is a larger model for solving complex problems, which is expected to be launched by the company between January and March. 

Krutrim is currently accessible to people in batches on simple signup, and the company will open it for all users by January 2024. Additionally, the developers can also access Krutrim APIs for innovation and experimentation starting February 2024. 

Aggarwal said, “Current AI models just can’t capture India’s culture, knowledge and aspirations, given our multicultural and multilingual context. We introduce Krutrim, a company with the sole vision of creating India’s AI for 1.4 billion Indians”. “India-first AI model should understand the uniqueness and the right cultural context. It needs to be trained in unique data sets specific to us. And on top of it all, it needs to be accessible to India, with India-first cost structures,” he added. 

Besides introducing its foundational AI model for generative AI applications, the company is also working on an AI computing stack in-house development & manufacturing of chips optimised for AI computing. The company has designed an architecture involving multiple chipsets to power different AI infrastructures, models, and applications.

Compared with GPT-4, which is OpenAI’s large language model to power ChatGPT, Krutrim performs faster in Indian languages by generating responses in less time, using less computing, the company said. In English, it outperforms Llama 2, which is Meta’s open-source large language model.