Summary
- Bengaluru’s Sarvam AI is targeting a trillion-parameter model, a scale that would put it in the same conversation as frontier labs abroad.
- The plan follows Sarvam’s $234 million Series B in June, which took it to unicorn status at a $1.5 billion valuation.
- At its Epoch conference today, Sarvam launched new 7B and 70B models first, treating the trillion-parameter model as the next big leap, not a finished product.
Sarvam AI wants to build a model with a trillion parameters. Say that out loud and it sounds like the kind of thing only OpenAI, Anthropic, or Google can pull off. A two-year-old startup out of Bengaluru is saying it anyway.
The company first signalled the plan back in May, when Sarvam co-founder circles confirmed on X that training would begin within nine months. Since then, the story has moved fast. Sarvam turned unicorn in June with a $234 million Series B led by HCLTech, hitting a $1.5 billion valuation, one of the largest AI funding rounds an Indian startup has ever closed. Part of that money is explicitly earmarked for what Sarvam calls its next frontier model, aimed at agentic, coding, and cybersecurity use cases.
Then, today, Sarvam hosted Epoch, its first flagship AI conference, in Bengaluru. The headline launch wasn’t the trillion-parameter model itself. It was Epoch Builder Edition, a new developer platform with fresh 7-billion and 70-billion parameter models trained on 2 trillion tokens, aimed at helping Indian developers and enterprises fine-tune and deploy their own AI systems. The bigger model is still ahead, not behind.
Why the scale matters, and why it’s still a plan
For context, Sarvam’s current flagship, Sarvam-105B, launched in February at the India AI Impact Summit, sits at 105 billion parameters. It uses a mixture-of-experts design that activates only around 10 billion parameters at a time, which is how a startup with a fraction of Big Tech’s compute budget can compete at all. Sarvam positioned that model against OpenAI’s GPT-OSS-120B and Alibaba’s Qwen-3-Next-80B.
A trillion-parameter model is roughly ten times bigger. For comparison, models in that range globally include Ant Group’s open-source Ling-1T and Moonshot AI’s Kimi K2. Frontier closed models from OpenAI and Anthropic are widely estimated to run even larger. So a trillion parameters gets Sarvam into serious company, but it’s not automatically a GPT-4 or Claude killer. Parameter count alone doesn’t decide who wins; training data, architecture, and how well the model actually performs on real tasks matter just as much.
It’s also worth being precise about where things stand. As of today, Sarvam has not announced that training has started, let alone finished. What’s confirmed: the intent, the funding tied to a “next frontier model,” and a public timeline that points to training beginning sometime around this year. Anyone reading this as a finished product announcement is reading ahead of the facts.
Read More: India’s First Homegrown Sovereign AI Model, Sarvam AI To Build
Why Sarvam is betting this big
Sarvam was founded in 2023 by Vivek Raghavan and Pratyush Kumar, both formerly with AI4Bharat at IIT Madras. It was the first company selected under India’s IndiaAI Mission, the government’s sovereign AI programme, and got access to Nvidia H100 GPUs in exchange for an equity stake. That government backing has been central to Sarvam’s pitch: build AI that’s trained in India, on Indian languages, run on Indian infrastructure.
The company says its platforms already handle over 2 million daily interactions and process close to 10 million API requests a day. Its voice agents have been used in a government farmer outreach programme reaching 17 million farmers, and in an insurance campaign covering 45 million policyholders. That’s meaningful usage for a company that was pre-revenue not long ago.
A model like Zoho or Krutrim scaling infrastructure is one story. A trillion-parameter foundation model is a different order of ambition entirely, and it puts real pressure on Sarvam to prove that Indian AI labs can play at frontier scale, not just build efficient smaller models for local use cases.
Read More: Krutrim’s Bold Leap: Unveiling Its Agentic AI Assistant Kruti
What this means for founders and the ecosystem
For Indian AI startups, Sarvam’s trajectory is a live case study in sequencing: build a working smaller model, get government backing, raise big once you’ve got usage numbers to show, and only then chase the very top of the scale ladder. It’s also a signal to enterprise customers evaluating sovereign AI options that Sarvam intends to stay in that conversation for the long run, not just ship one flagship model and coast.
Whether the trillion-parameter model actually lands, and whether it can compete on more than just size, remains to be seen. But the ambition itself says something about how far India’s AI ecosystem has come in under three years.
What do you think, can an Indian startup genuinely compete at the same parameter scale as OpenAI and Anthropic, or is this more about signalling than substance? Drop your take in the comments, and keep following Startup INDIAX for how this story develops.
FAQs
What is Sarvam AI’s trillion-parameter model?
It’s a planned large language model that Sarvam AI says it intends to train, targeting roughly a trillion parameters, a scale meant to put it in direct competition with frontier models from OpenAI and Anthropic.
Has Sarvam actually started training the trillion-parameter model?
As of this article, Sarvam hasn’t confirmed training has begun. The plan was first reported in May 2026, with funding from its June Series B partly earmarked for this “next frontier model.”
How is this different from Sarvam-105B?
Sarvam-105B, launched in February 2026, has 105 billion parameters. A trillion-parameter model would be roughly ten times larger, requiring far more compute and training data.
Why did Sarvam launch Epoch Builder Edition instead of the big model?
Sarvam is building up in stages. Epoch Builder Edition, launched July 30, gives developers new 7B and 70B models to build with now, while the trillion-parameter model remains a longer-term goal.
Who funds Sarvam AI?
Sarvam is backed by Lightspeed Venture Partners, Peak XV Partners, and Khosla Ventures from earlier rounds, with HCLTech and Bessemer Venture Partners joining as lead investors in its $234 million Series B.