Bhavish Aggarwal's Krutrim launches an AI chatbot to take on ChatGPT, Gemini

Krutrim will be powered by the company's multilingual large language models (LLM), also called Krutrim.

Krutrim, an artificial intelligence startup launched by Ola founder and chairman Bhavish Aggarwal, on February 26 rolled out an AI chatbot in public beta, that works similar to OpenAI's ChatGPT and Google's Gemini.

The launch comes a month after Krutrim disclosed a $50 million financing at $1 billion valuation, to became the country's first startup unicorn in 2024. The company mentioned that it is the first AI unicorn in the country.

Matrix Partners India, an early backer of Aggarwal's other ventures Ola Cabs and the IPO-bound Ola Electric, led the funding round.

The chatbot, which bears the same name as the company, was announced in December. It is the firm's first product, which will be powered by its multilingual large language models (LLM), also called Krutrim.

"This is a start for us and our first generation product. Lots more to come and this will also improve significantly as we build on this base" Aggarwal said in a post on X, the social media platform previously known as Twitter.

He said the chatbot will assist consumers in over 10 Indian languages including Hindi, Tamil, Bengali, Marathi, Kannada, Gujarati, besides Hinglish (a mix of Hindi and English) and English.

"While some hallucinations will be there but much lower for Indian contexts than other global platforms. And we will be working overtime to find and fix" Aggarwal said.

That said, we observed that Krutrim currently answers questions only in English and Hindi at present, although it does understand questions from several Indian languages.

A screenshot of Krutrim chatbot. It currently provides answers in English and Hindi

One can rate the responses, copy them to use in any other application or even regenerate them.

Krutrim's full-stack ambitions

Krutrim, which means "artificial" in Sanskrit, aims to focus on building the entire AI computing stack.

Trained by a team of computer scientists, based in Bengaluru and San Francisco, the company's LLM will be available in two different sizes: a base model named Krutrim base, and a larger, more complex model called Krutrim Pro.

Krutrim Pro, launching in Q4 FY 2024, will be multimodal in nature, which means it can understand and work with different formats, including text, audio, image, and video, at the same time. It will also have larger knowledge, advanced problem-solving and task execution capabilities.

Krutrim claims that these models have been trained on 2 trillion tokens of data with the largest representation of Indian data to understand the nuances of Indian culture and languages, without disclosing any further details. It can understand 22 Indian languages and generate text in 10 Indian languages, the company said.

At the launch event in December, Krutrim claimed that its AI models perform better than many open-source LLMs trained with similar amounts of data on a range of industry benchmarks. It also claimed to outperform OpenAI's GPT-4 in Indic language performance.

Krutrim will operate as a separate company within the Ola Group, similar to Ola Electric and its cab business. However, it does not use data from these businesses to train its AI models, Aggarwal said during the launch event.

Krutrim plans to offer its AI models as an API for enterprises and developers, seeking to create AI applications, on an "India-first cost structure", Aggarwal said.

In addition, Krutrim plans to build its own silicon and design its own chips. The company is also working on AI infrastructure to develop indigenous data centers and eventually, server-computing, edge-computing and super-computers. The production is scheduled for mid 2024 for prototypes and a roll out production roadmap by the end of 2025.

Ola also plans to integrate Krutrim across its group companies by March 2024, leveraging the technology for sales, service, support, and other key processes.