Sarvam

Sarvam made a bet most Indian software companies never had the compute to make: build the models, not just the apps on top of them. Founded in August 2023 by Vivek Raghavan and Pratyush Kumar, both out of the AI4Bharat lab at IIT Madras, it set out to build a full-stack sovereign AI platform for a country that speaks hundreds of languages and where English is a minority tongue.

That bet got official in April 2025, when the government selected Sarvam under the IndiaAI Mission to build the country's first sovereign large language model on state-backed compute. By June 2026 the company was a unicorn, and in February 2026 it open-sourced two frontier models trained entirely on that mission compute. The work runs out of Bengaluru, with a second team in Delhi.

Sarvam builds the whole stack: the foundational models, the infrastructure they run on, and the enterprise applications on top. The models are tuned for Indian languages and voice first, so a bank in Bengaluru or a hospital in Delhi can run AI that understands Hindi, Tamil, Telugu, and a dozen more without routing every request through a US or Chinese provider.

The pitch to enterprises and government is sovereignty as much as capability. Everything is developed and operated in India, with deployment options that reach all the way to air-gapped, on-premise environments for regulated buyers who cannot send data anywhere.

Most frontier models are trained on English and a handful of high-resource languages, which leaves India's speakers on the outside of the AI economy. Sarvam's answer is models that reason, speak, and transcribe across roughly 12 to 23 Indian languages, plus the plumbing to run them at population scale.

The second problem is control. Government and BFSI buyers cannot hand sensitive data to a foreign API, so Sarvam's edge is that the model, the compute, and the data all stay inside the country, with private, hybrid, and air-gapped deployments for the strictest cases.

Sarvam sells to enterprises, government, and developers, but the center of gravity is regulated, high-volume Indian institutions. Banks and insurers want multilingual customer engagement that works over voice and WhatsApp; government bodies want AI they can run on sovereign infrastructure; developers want Indic-language APIs they can build on.

The named customers back this up: Tata Capital runs Sarvam's Samvaad voice platform across consumer loans in English plus ten Indian languages, and SBI Life put a Sarvam WhatsApp assistant in front of its entire 350,000-plus agent force in eleven languages.

Sarvam's catalog runs the full stack, from raw models to finished enterprise agents. At the base are the speech and language models, Saaras for speech-to-text, Bulbul for text-to-speech, translation across many languages, and the vision model Akshar for turning documents into structured data, all served through Sarvam Cloud.

On top sit the applications: Samvaad for building voice and chat agents, Arya for AI-at-work with forward-deployed engineers embedded in customer teams, and Studio, a media platform for AI dubbing, live translation, and voice cloning across 12-plus languages. In February 2026 Sarvam open-sourced two of its foundation models, the 32B-parameter Sarvam 30B and the 106B-parameter Sarvam 105B, both mixture-of-experts designs.

Sarvam makes money the way an AI infrastructure company does: enterprise and government contracts for its models and applications, plus a developer API business through Sarvam Cloud. The forward-deployed engineering motion, engineers embedded directly with customers, is central to how deals land and expand in verticals like banking, insurance, and government.

The open-source models play the Mistral-style hybrid game: give the weights away to build ecosystem and trust, then monetize scale through hosted APIs, enterprise deployment, and custom agent work. Sarvam does not publish standard pricing, so the commercial terms are contract-based.

Sarvam competes on two fronts at once: global frontier labs like OpenAI, Anthropic, and Google whose models dominate general capability, and the wave of Indian sovereign-AI efforts the government is funding. Its wedge is not raw benchmark supremacy, it is being India-native by construction, the only player with models, compute, and data all inside the country.

The company claims its 105B model matches or beats most open and closed models in its class on general benchmarks and pulls clearly ahead on Indian-language ones. Whether that translates into a durable moat depends on developer adoption before the global labs ship strong multilingual models of their own.

The traction story is about volume in the real world. SBI Life's deployment reaches an agent network of more than 350,000 distributors and is aimed at over 80 million customers, which the company frames as the largest enterprise WhatsApp use case in India. Tata Capital runs Sarvam's voice AI across consumer-loan customer interactions in eleven languages.

The customer roster, Tata Capital, SBI Life, CRED, IDFC, and LIC, reads as a who's-who of Indian financial services, and the government's IndiaAI selection is its own kind of proof: the state chose Sarvam to build the model it will stand behind.

The praise is loud and directional: Sarvam is widely seen as India's most credible shot at a sovereign frontier model, and the decision to open-source the 30B and 105B weights won goodwill as a real bet on open AI rather than a walled garden. The Indian-language benchmark results are the part reviewers take most seriously.

The criticism is practical. Developers flagged missing GGUF formats and thin support for common inference frameworks like vLLM, which makes the open models hard to actually run locally. The bigger open question is economic: training 100B-plus models burns cash, and Sarvam still has to build a developer ecosystem deep enough to justify it.

Sarvam raised roughly $41 million across seed and Series A in December 2023, led by Lightspeed with Peak XV and Khosla Ventures, the largest AI round in India at the time for a company barely five months old. The real jump came later: in June 2026 it announced $234 million as the first close of a $300 million Series B at a $1.5 billion post-money valuation, tipping it into unicorn territory.

HCLTech led that round with $150 million as a strategic investor, joined by Bessemer Venture Partners and returning backers Khosla Ventures and Peak XV. The money is earmarked for training the next frontier model, aimed at agentic, coding, and cybersecurity use cases, and for the compute to expand its forward-deployed motion.

Sarvam frames itself as a small, high-talent-density team that moves fast and holds a very high bar, and the careers language leans hard on ownership: high agency, high impact from day one, AI-first in how they build and think. The founding DNA is research, Pratyush Kumar and Vivek Raghavan came from AI4Bharat and India's open-source Indic-language and digital-public-infrastructure work, but the culture prizes shipping things that run in production, not demos.

The forward-deployed engineering model is a cultural tell as much as a go-to-market one: engineers sit inside customer teams and own delivery end to end. It is an on-site-first company, with roles based in Bengaluru and a Delhi team for its strategic and defence-facing Chanakya vertical.

For Sarvam, compliance is the product pitch, not a footnote. The company reports SOC 2 Type II, ISO 27001, and DPDP compliance, and its whole reason for existing is data control: everything built and operated in India, under Indian data-protection law.

That posture is what unlocks its buyers. Deployment stretches from managed cloud to private, on-premise, hybrid, and fully air-gapped environments, so banks, insurers, and government bodies can run the models without data ever leaving their own walls.