The company, which built the model with under $5,00,000, has launched an upgraded version – Project Indus 2.0.
Tech Mahindra’s Project Indus, the maiden attempt by India Inc, to build a cost-effective large language model (LLM), is looking to expand its scope by launching a similar LLM for Bengali audiences. It is also working on an agentic application in Hindi and its dialects. (An agentic application, often driven by AI such as LLMs, is a software system that can independently work towards goals for a user.)
The LLM, which can be accessed on the Project Indus website, witnessed 8,000 downloads so far. “The idea is to make the model available for the larger community within India. Indus also became the inspiration for Sahabat-AI, a Bahasa-based sovereign LLM built for Indonesia,” Nikhil Malhotra, Chief Innovation Officer & Global Head of AI and Emerging Technologies, Tech Mahindra, said.
“We are expanding Hindi language and dialects for our agentic solutions. Additionally, we are expanding our database. The next language we will focus on is Bengali,” he said.
The company, which built the model with under $5,00,000, has launched an upgraded version – Project Indus 2.0.
“Project Indus is undergoing continuous advancements to stay at the forefront of innovation. More data is being added for dialects through supervised fine-tuning,” Malhotra said.
An LLM on a chip
The company is working on a project where it wants to build a small model with 1.2 billion parameters, which would make it ideal for deployment on chips.
“This would be useful in edge scenarios, such as farms, factories, and the Indian automotive industry. It is also gaining traction across industries, particularly in automotive, agriculture, and finance,” he said.
With support for Hindi and 37 dialects, Project Indus provides voice and chatbot interaction. You can download it from Hugging Face’s website or use it online in India.
Independent study
“An independent study conducted by Assam Kaziranga University in Jorhat compared various models and demonstrated that Indus outperformed many of its counterparts in tokenisation efficiency,” Malhotra said.
“The speed of tokenisation is influenced by the number of parameters and the model’s size. It is safe to say that Indus has shown it is possible to outperform even larger models by using optimised techniques,” he added.
“We are also leveraging LLMs to enhance our services for customers and improve delivery. We are utilising OpenAI models to fine-tune our LLMs. The usage varies depending on the specific use case, which is determined through our research to compare the models and select the appropriate one,” he said.
“The Indus LLM has many applications across India. For example, Tech Mahindra group companies such as Mahindra Finance are looking to potentially use it to support low-income families with financial services in their dialects, while students benefit from chat-based AI features,” Intel has said in its assessment of Tech Mahindra’s LLM project.
Published on April 7, 2025
