Ola’s Bhavish Aggarwal launches India-focused AI base model Krutrim

Ride-hailing app Ola’s Bhavish Aggarwal has launched Krutrim, an artificial intelligence base model focusing on Indic languages, cultural scenarios, and settings. Touting it to be the first of its kind, Aggarwal said that Krutrim could understand India’s ‘uniqueness and right cultural context’. Krutrim is already being used by different teams in the company for activities like customer support, sales, and human resources tasks, among other processes.

To be sure, Aggarwal launched his AI venture called Krutrim Si Designs to compete with the likes of OpenAI in April. In the past, Aggarwal had spoken about the need for India’s own foundational models and libraries. 

Krutrim is trained on 2 trillion tokens. It can understand queries in 20 languages and generate responses in 10 of them including Marathi Hindi, Bengali, Tamil, Kannada, Telugu, and Odia. The spokesperson said that the model has been tested against several benchmarks on parameters like Indian English, reasoning and logic, among others, and was found to surpass in performance against similarly sized model Llama 2 from Meta. The model will open for use by next month. During the demo, Krutrim was shown to perform tasks such as writing poems in Tamil and Bengali, developing a small coding snippet, and generating a story.

A multimodal version of Krutrim, called Krutrim Pro which can process audio and image inputs, will be available next quarter. “(Krutrim) Pro will have more sophisticated problem-solving and task execution capabilities and will power several future products and applications. This is of course the start. We will continue to build other AI models across text, voice, and vision and beyond LLMs,” said Ravi Jain, head of strategy at Ola Tech Solutions. 

Beyond the Krutrim AI model, the team is also currently working on building its own novel chip architecture. This system on package (SoP) is energy efficient and cost-effective in India’s context, the team said. As the name suggests, SoP is a technique of bundling two or more integrated circuits in the same package. In the future, the team plans on scaling up the SoP to build clusters and ultimately a supercomputer. The first prototype is expected to be out by next month.

Lastly, Krutrim’s team is also working on another layer that focuses on making data centers more energy-efficient, especially for AI workloads. “We are working ground up on its cooling technology, the rack technology powering up to 200 kilowatts, and we want to build networks which will be up 200,000 GPUs in the cluster. We are working on AI servers and the whole data center design which can power the AI for the future,” the spokesperson said.

To be sure, Sarvam AI, the startup that emerged from stealth this week and secured funding of $41 million, launched OpenHathi, a Hindi-based LLM. Unlike Krutrim which is a base model, OpenHathi has been fine-tuned on Meta’s Llama 2 model.