Bengaluru: Sarvam AI, the India-based AI platform that turned a unicorn last month, has announced plans to build a trillion-parameter model, which will take it closer in capabilities to the world’s frontier models like ChatGPT, Gemini and Claude. Towards this, the company said it is also establishing an office in the US to attract the best Indian-origin AI researchers based there.Speaking at the company’s flagship developer meet in Bengaluru on Thursday, co-founder Pratyush Kumar said the new model is aimed at competing with the frontier models in coding, cybersecurity, scientific research, and simulation.He said the current 100-billion-parameter model is already strong for voice AI, conversational systems, English speech recognition, dictation, and even some complex tasks such as simulating physics lessons. He said Sarvam has already processed 325 million minutes of voice calling, and that in one benchmark the model was only 4 to 5 percentage points behind a much larger model.In comparison with global models, he said Sarvam’s voice systems were 4 to 5 times cheaper. And this is not just for Indian languages. Even English speech recognition, he said, is state-of-the-art, and way cheaper.Kumar and co-founder Vivek Raghavan argued that their efforts, including the move to build a trillion-parameter model, are to ensure India does not just consume AI, it also manufactures intelligence. “If India keeps importing tokens (the unit of generation produced by large language models) from global leaders such as OpenAI or Anthropic, it risks paying not just in money but also in data (because India is continuously providing data to those companies to further train their models),” Kumar said, adding, “India should be producing the AI tokens it consumes.”Raghavan said Sarvam is serious about building a “real token factory in India”. “This is important because if we can serve tokens from India, that’s the first step to becoming sovereign,” he said. In a deglobalising world, there are worries that countries like the US and China could stop others’ access to their frontier AI models.Kumar also rejected the idea that India should accept weaker products simply because they are homegrown. “Sovereignty should not be a tax,” he said, insisting that Indian AI must be globally competitive and benchmarked against the best in the world.He said that is eminently possible, just the way China has done. This is the reason Sarvam is establishing a US office. “We’re getting researchers from the US to work with us, we’ll identify the research problems we are going after,” he said. The company has already made one appointment – of renowned AI researcher Devendra Chaplot, a founding member of Mistral AI and Thinking Machine Labs. California-based Chaplot will be an advisor.Sarvam on Thursday also announced an India-hosted inference platform, and a slate of new models and products such as AI-powered sunglasses under its wearables vertical Kaze; a Python SDK that lets developers submit datasets to Sarvam for training; its latest multilingual text-to-speech model Bulbul V4; its latest vision-language model Sarvam Vision 2.0; and Indus, an agentic AI platform which brings together work, voice, and coding agents.Kumar said the company’s biggest business traction today is in BFSI – banks and financial institutions use Sarvam’s stack for voice and agent-based workflows. He pointed to an AI agent deployed for SBI Life’s 300,000 sales people, which he said made them far more effective. Sarvam is also seeing interest from digital-native firms, defence and intelligence agencies, and the public sector.
