Sarvam AI Is Building a Trillion-Parameter Model — Entirely in India

Sarvam AI just made the boldest claim any Indian AI company has made so far: it’s building a model with more than a trillion parameters, trained from scratch, on Indian soil.

The Bengaluru-based startup dropped this at Epoch 2026, its first-ever developer conference, held in the city on July 30, 2026. It wasn’t a one-line teaser buried in a slide — it came with a working inference platform, four new product upgrades, a Silicon Valley office, and a marquee hire, all announced in the same afternoon.

If you’ve been tracking India’s “sovereign AI” story since the IndiaAI Mission kicked off, this is the moment that story stopped being about GPU allocations and government tenders, and started being about an Indian company trying to sit at the same table as OpenAI, Google, and Anthropic.

Here’s everything that was actually confirmed, what’s still unverified, and why it matters if you’re building, investing, or writing code in India right now.

What Sarvam AI Announced At Epoch 2026

Sarvam cofounder Pratyush Kumar opened the event with the headline announcement: the company is building a model with over a trillion parameters, aimed squarely at coding, cybersecurity, scientific research, and simulation — the categories where frontier labs currently dominate. The model is being built “right here in India,” Kumar said, from scratch rather than fine-tuned on an existing open-weight base.

One important caveat, and this is where a lot of hype cycles fall apart: Sarvam did not share a launch timeline. No date, no quarter, no year. For now, this is a stated direction, not a shipped product. Worth remembering the next time a headline treats it as done.

The Real Product Launch: Sarvam Inference

While the trillion-parameter model grabbed headlines, the more immediately usable announcement was Sarvam Inference — a new platform that serves frontier open-source models using infrastructure hosted entirely within India.

At launch, Sarvam Inference supports three models: Sarvam’s own 105-billion-parameter model, plus the open-weight GLM 5.2 and Gemma 4. The pitch is simple — Indian developers and enterprises can call frontier-grade models through an India-hosted endpoint instead of routing every token through a US or EU cloud region.

Cofounder Vivek Raghavan framed this around a concept he called “token sovereignty” — the idea that a larger share of the AI tokens Indian companies consume every day should run through domestic infrastructure, not overseas servers. That phrase is likely to become a talking point well beyond this one product launch, especially with government and BFSI clients who have hard data-residency requirements.

Sarvam also said it has built an agentic optimisation layer that automatically tunes model performance, claiming up to a 15x improvement in inference speed on certain models. That’s a strong number — and one we’d want to see reproduced independently before treating it as gospel.

Voice, Vision And The Rest Of The Stack

Sarvam has spent the better part of 2026 arguing that owning a single chatbot isn’t enough — it wants the entire AI stack, from the model down to the voice interface. Epoch 2026 added four upgrades to that stack:

A number of additional specifics — including exact API pricing, a claimed 5.5x cost advantage over rival “mini” models, and usage figures for Sarvam’s Indus agent platform — were shared by attendees on X during the event rather than confirmed in Sarvam’s own release. Those numbers are plausible given Sarvam’s pricing history, but until the company publishes them on its own blog or API docs, treat them as provisional.

Why This Actually Matters For Indian Developers And Startups

Sarvam has quietly become the closest thing India has to a national AI champion. It’s one of 12 companies backed under the ₹10,372-crore IndiaAI Mission, it was the breakout name at the India AI Impact Summit in February, and the government is reportedly on track to hold a small equity stake in the company through its compute-subsidy arrangement.

What changed today is the scope of ambition. Building a sovereign 105-billion-parameter model for Indian languages is one kind of project. Announcing a trillion-plus-parameter frontier model, an India-hosted inference layer competing with global cloud providers, and a Silicon Valley office — on the same day — is a different kind of statement. It puts Sarvam in direct rhetorical competition with the same labs India has spent the last year trying to reduce its dependence on.

For Indian developers specifically, Sarvam Inference is the more immediately practical story. If it holds up, it’s a way to run frontier-class open models without your data leaving the country — a genuine gap in India’s current AI tooling landscape.

The Reality Check

Three things worth flagging before this becomes another “India beats Big Tech” headline:

Bottom Line

Sarvam AI didn’t just launch a product at Epoch 2026 — it staked a claim to being India’s frontier-AI company, not just its sovereign-language-model company. The trillion-parameter model is a promise with no delivery date. Sarvam Inference is real, live, and arguably the more important announcement for anyone building in India today. Expect both to be tested hard over the next two quarters — by developers, by competitors, and by the government that’s partly funding this bet.

FAQs