Sarvam AI Announces 1 Trillion-Parameter Model, Hires xAI Researcher, and Launches Full-Stack AI Platform to Challenge OpenAI and Anthropic

Sarvam AI has just had its biggest day. At its inaugural Epoch 2026 conference in Bengaluru, the sovereign AI startup announced it is building a one trillion-parameter foundation model from scratch in India - placing it squarely in competition with OpenAI, Anthropic, and Google on the global AI frontier, while simultaneously launching a full-stack AI platform designed to make it a comprehensive infrastructure provider for Indian enterprises, developers, and government institutions.

The company outlined an ambitious roadmap to become a full-stack AI platform spanning infrastructure, foundation models, enterprise agents, and developer tools - signalling its intent to compete not only with global model makers such as OpenAI and Anthropic but also with hyperscale cloud providers.

The Trillion-Parameter Bet

Sarvam said it is building a one-trillion-parameter foundation model from scratch in India, designed for coding, cybersecurity, scientific research, and simulations - with the explicit goal of reducing India's dependence on foreign AI models. The company committed to building the trillion-parameter model within the next six months, arguing that India must develop frontier AI capabilities rather than rely solely on overseas providers.

The ambition is significant. OpenAI's GPT-4 was widely estimated to have around 1.8 trillion parameters. A trillion-parameter model built and trained entirely in India - on Indian infrastructure, with Indian data - would represent an entirely different category of achievement from the fine-tuned or smaller open-source models that most Indian AI labs have focused on.

A Full Suite of New Products

Alongside the flagship trillion-parameter announcement, Sarvam unveiled a comprehensive stack of upgraded and new products.

The upgraded Sarvam 105B model now handles complex tasks, supports voice, and is priced at $0.80 per million blended tokens - 5.5 times cheaper than GPT-5.4 Mini, with an additional 30% price cut announced at the event. Sarvam also unveiled Sarvam Vision 2.0, an AI model that can read Indian handwriting, scan tables, read forms, and automatically identify details like names, addresses, and IDs.

On the voice side, Bulbul V4 turns text into natural-sounding speech with emotions including laughter, excitement, and emphasis, while Saras V4 speech-to-text now ranks among the world's best for English while supporting several Indian languages including Odia, Sanskrit, and Manipuri. A multi-speaker version of Saras V4 can separate and transcribe multiple speakers simultaneously.

Sarvam also launched Sarvam Inference - allowing developers to run leading open AI models on servers located in India, providing faster responses and keeping data within the country.

The xAI Hire

Sarvam brought onboard AI researcher Devendra Singh Chaplot, who most recently worked at Elon Musk's xAI, as a part-time advisor - marking a significant talent gain for the Indian company. Chaplot has also been part of the founding teams at Mistral AI and Mira Murati's Thinking Machines Lab. His arrival signals that Sarvam is now attracting researchers who have operated at the bleeding edge of global AI development.

Platform, Partnerships, and Global Expansion

The new Epoch Builder Edition platform is aimed at banks, government departments, startups, and research institutions that want to build AI applications using their own data. Sarvam also announced partnerships with three IITs and two state governments for pilot projects in education and public service delivery, with Epoch Builder Edition set to enter private preview in August 2026.

Sarvam has also opened a San Francisco office, signalling that while its mission is sovereign AI for India, its commercial ambitions are global.

The Bigger Picture

The strategy marks a significant shift for the Bengaluru-based startup, which says the next phase of AI adoption in India will be driven less by standalone models and more by complete AI stacks that enterprises can deploy within the country.

Backed by Lightspeed and Khosla Ventures with over $50 million raised, Sarvam is currently valued at around $200 million - tiny next to OpenAI's $500 billion. But the gap in ambition is closing faster than the gap in valuation. A trillion-parameter model built and priced for India, trained on Indian infrastructure under the IndiaAI Mission, with a full-stack platform underneath it, is a different kind of company than the one that debuted a 2-billion parameter model in October 2024.