Bengaluru firm unveils two AI language models

A few months after the union government announced that India would create its own Large Language Model (LLM) such as ChatGPT and DeepSeek, the Bengaluru-based Sarvam AI unveiled two language models at the AI Impact Summit here on Wednesday (February 18, 2026). Named Vikram, the models will be open source, said Pratyush Kumar, co-founder at Sarvam.

India AI Summit 2026 Day 3 Highlights

Mr. Kumar presented them at a keynote address during the summit’s AI Research Symposium. Both models, he said, had beat other comparable models from around the world at industry benchmarks. 

Training a large language model is a computation-and-skill-intensive process, sometimes entailing millions of dollars worth of work put in by graphics processing units (GPUs) working in concert at data centres. The Ministry of Electronics and Information Technology has instead emphasised inference (the act of actually querying a trained model) as opposed to training. However, in the period after the launch of the Chinese LLM DeepSeek R1, IT Minister Ashwini Vaishnaw said an Indian firm would also be able to train an LLM.

Sarvam had help from private investors such as Peak XV and Khosla Ventures, who collectively put in upwards of $50 million into the firm. It also had subsidised access to GPUs under the IndiaAI Mission’s common compute programme. 

A key aim of the models has been to be better at Indian languages, which don’t work very well on most AI models owing to the relative scarcity of text content available in Indian languages that could be used to train models. Mr. Kumar did a demonstration, translating English content into Indian languages.

The launch of the much-anticipated models – a 35-billion parameter model and a 105-billion parameter model – is being seen as a milestone for AI development in India.

These are yet to be made available for the public. A message on the Sarvam website says that a chat feature will be available “soon”.