Tech Mahindra's indigenous Gen AI project Indus now in Beta mode: CP Gurnani

CP Gurnani, the outgoing chief of IT services major Tech Mahindra, tweeted on social media platform X (formerly Twitter) that the generative artificial intelligence project Indus has been ‘successfully completed’ and has been already launched in beta phase within the company by its innovation unit called the Makers Lab. “The model is probably only one in the world which has all Hindi tokens and has been trained grounds up. It will set the stage for the years to come as our prowess in deep tech,” said Gurnani.

Project Indus was announced in August 2023. The company had then said that with this project, it aims to construct an indigenous large language model (LLM) designed to converse in various Indic languages, inspired by the government-supported AI translation tool Bhashini. As per the latest update from Gurnani, Project Indus currently has a Hindi LLM with 539 million parameters and 10 billion tokens. Tokens can be defined as the units of text that a model can read or generate.

"The model has a user-friendly interface to demonstrate its capability, and we have released it in a beta version for our employees. We trained the model on almost 114 GB (GigaByte) of Wikipedia, news data, and book corpora, which we collected over the first three months and then cleaned, annotated, and translated," Nikhil Malhotra, Global Head, Makers Lab, Tech Mahindra told TechCircle in a written response. Phase 1 of the project serves the Hindi language and its 37 dialects; the team will now focus on fine-tuning it for downstream tasks, he added.

As LLMs continue to become more efficient and effective, there has been a call for building localised versions that can better cater to region-specific needs. Currently, a vast majority of these models cater to English-speaking audiences.

Last week, Ola’s Bhavish Aggarwal launched Krutrim, an AI 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 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 team has yet to release details on the technical specifications or the financial cost incurred in building such a massive model. To set the context, ChatGPT-builder OpenAI spent an estimated $100 million and used more than 30,000 GPUs to build generative pre-trained transformer (GPT) model.

Further,  Sarvam AI, the startup that emerged from stealth recently 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.