Sarvam to build a trillion-parameter model from scratch, unveils 12 products at Epoch 2026

India's boldest model bet, made in Bengaluru

At its first developer conference, Epoch 2026, held in Bengaluru on 30 July 2026, Sarvam AI confirmed it is building a foundation model with more than one trillion parameters from scratch in India — a scale that would put it in the same weight class as frontier systems from OpenAI, Google and Anthropic. Co-founder Pratyush Kumar told the audience the model would go live within roughly six months, and that it is being trained specifically to be competitive in coding, cybersecurity, simulation and scientific research rather than only in chat.

For a company that was, barely a year ago, best known for compact Indian-language models, the announcement is a deliberate step up in ambition — and a statement that a home-grown team intends to build at the frontier, not merely fine-tune other people's weights.

Twelve launches, and a pitch on price

The trillion-parameter roadmap headlined a broader product blitz. Sarvam used Epoch to roll out around a dozen products, including Bulbul V4, an expressive text-to-speech system; Sarvam Vision 2.0 for document understanding; and the Epoch Builder Edition, a platform aimed at developers and enterprises building, fine-tuning and deploying language models for Indian languages and use cases.

Cost was central to the pitch. Sarvam said its coming flagship would be priced roughly five times cheaper than global rivals. It illustrated the argument with its existing 105-billion-parameter model, which it said it serves at about $0.80 per million blended tokens — against a claimed $4.50 for OpenAI's GPT-5.4 Mini and $9 for Google's Gemini 3.5 Flash. If those numbers hold in practice, the company argues, Indian developers could run capable models at a fraction of the cost of importing inference from abroad.

"Token sovereignty"

Alongside the models, Sarvam announced an India-hosted inference service that lets developers run leading AI systems on domestic servers. Co-founder Vivek Raghavan framed the move as a push toward what he called "token sovereignty" — serving a larger share of the AI compute consumed within India from infrastructure physically located in the country, rather than routing queries and data through overseas clouds.

That framing dovetails with the government's own priorities. Sarvam was selected in 2025 under the IndiaAI Mission to build a sovereign large language model, and the mission has been assembling a shared GPU pool — tens of thousands of accelerators offered to startups at subsidised hourly rates — precisely to make training and serving large models in India economically viable.

Can a startup train at the frontier?

Building a trillion-parameter model from scratch is expensive and technically punishing. It demands enormous, reliable compute; high-quality, well-curated training data; and a team that can wrangle distributed training at scale without burning months on failed runs. Sarvam, which reached unicorn status at roughly a $1.5 billion valuation, is betting that frugal engineering and access to subsidised national compute can close part of the gap with far better-funded Western labs.

Sceptics will note that promised parameter counts and headline prices are easier to announce than to sustain, and that a six-month timeline for a from-scratch frontier model is aggressive. Independent benchmarks — not launch-day claims — will decide whether Sarvam's model is genuinely competitive in the hard domains it is targeting.

Why it matters

Whatever the outcome, Epoch 2026 reframes the Indian AI conversation. For years the debate was whether India should build foundational models at all or simply adapt open-weight systems. Sarvam's answer is unambiguous: build, at scale, at home, and compete on price. If it delivers even part of that roadmap, it strengthens the case that sovereign, cost-efficient AI infrastructure is within reach for a country that has the developers and the data but has lacked frontier-scale models of its own.

Sources

• Inc42 — Sarvam to build trillion-plus AI model in India, launches inference service: https://inc42.com/buzz/sarvam-to-build-trillion-plus-ai-model-in-india-launches-inference-service/
• Free Press Journal — Sarvam bets big on India's AI future with trillion-parameter model: https://www.freepressjournal.in/tech/sarvam-bets-big-on-indias-ai-future-with-trillion-parameter-model-will-be-priced-five-times-cheaper-than-global-rivals
• The Hans India — Sarvam to host Epoch AI event on July 30: https://www.thehansindia.com/tech/sarvam-to-host-epoch-ai-event-on-july-30-model-updates-and-product-launches-1102384
• Trak.in — Made in India AI Sarvam to build 1-trillion-parameter model: https://trak.in/stories/made-in-india-ai-sarvam-to-build-1-trillion-parameter-model/