Sarvam AI: Pioneering India’s Sovereign AI Future

Lisa Ernst · 13.02.2026 · Artificial Intelligence · 8 min

India is racing to build “sovereign AI” — not as a buzzword, but as a practical strategy: keep data, models and compute under national trust boundaries while serving a population that speaks hundreds of languages and dialects. The country already runs some of the world’s largest digital public systems, from Aadhaar to payments at population scale. The next frontier is making AI work for everyone, not just for English-first users.

That’s where Sarvam AI comes in: a Bengaluru startup founded in August 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar (Sarvam “About Us”). Sarvam positions itself as a builder of India-first foundation models — speech, vision/document intelligence, and large language models — optimized for real-world Indian usage (including code-mixing, where people switch languages mid-sentence).

Source: An illustration shows neural networks connecting various global elements, symbolizing the worldwide reach and impact of AI innovation.

Quick Summary: What is Sarvam AI?

Sarvam AI is an Indian AI company focused on building sovereign, Indic-first AI — models that understand Indian languages, accents, documents and “messy” real-world audio.

• Founded: August 2023 (Bengaluru) by Dr. Vivek Raghavan and Dr. Pratyush Kumar (Sarvam).
• Government selection: Chosen under the IndiaAI Mission to build India’s sovereign LLM (announced April 26–27, 2025; Sarvam’s own announcement: “India’s sovereign LLM”).
• Key model areas: Speech recognition (ASR), document intelligence/OCR, and text-to-speech (TTS).
• Funding: Reported around $53M+ raised (e.g., coverage and trackers; Sarvam announced a $41M Series A on Dec 7, 2023: Series A announcement).

Sarvam AI’s Sovereign Ambition: Why India Cares

“Sovereign AI” usually means three things: (1) data governance (where sensitive citizen and enterprise data lives), (2) model control (what the model is trained on and how it behaves), and (3) compute availability (the GPUs and infrastructure to build/serve models). India’s own language stack matters here — not only for inclusion, but also for public-service reliability. A government helpline or a farmer assistant cannot fail because a model doesn’t understand an accent, a regional script, or mixed Hindi-English queries.

Sarvam’s approach aligns with national programs like INDIAai / IndiaAI and language initiatives such as BHASHINI, which aim to remove language barriers for digital services. Sarvam’s bet is that the “last mile” of AI in India is not a novelty chatbot — it’s voice, documents, and workflow automation across Indian languages.

The Visionary Founders

Source: coe-iot.com

Vivek Raghavan’s background includes building digital public infrastructure at population scale — a relevant skillset when your AI needs to work reliably for millions of users.

Dr. Vivek Raghavan (IIT Delhi; PhD Carnegie Mellon) has a long track record in large-scale systems and digital infrastructure. He is often associated with work around India’s national platforms and public-interest technology, a theme Sarvam continues by pushing “AI as infrastructure” rather than “AI as a toy.”

Source: businesstoday.in

Pratyush Kumar helped build India’s open-source momentum in language AI. That matters when you need datasets, evaluation and community trust.

Dr. Pratyush Kumar (IIT Bombay) is closely linked to India’s open-source language AI ecosystem, including AI4Bharat (IIT Madras), which publishes datasets and models for Indian languages. Sarvam benefits from that ecosystem — because for Indian language AI, the “secret sauce” is often the data: scripts, transliteration, noisy speech, and domain documents.

Breaking Benchmarks (and What They Actually Mean)

Headlines often say “model X beats ChatGPT.” The important detail is where and how. Sarvam’s recent momentum is strongest on India-centric speech and document tasks, where global models sometimes struggle due to accents, code-mixing, and complex local document formats.

Sarvam Audio (Speech-to-Text for Indian languages)

In early 2026 coverage, Sarvam claimed its speech recognition system (Sarvam Audio) achieved lower Word Error Rate (WER) on the IndicVoices dataset than systems like GPT-4o Transcribe and Gemini variants. WER is the standard metric in speech recognition: lower is better — fewer substitutions, deletions, and insertions in the transcript. The practical implication: better outcomes for call centers, citizen helplines, healthcare intake, and any “voice-first” workflow where English is not the default. (Coverage example: Business Today.)

One especially India-relevant capability is handling code-mixing — people routinely blend Hindi/English, Tamil/English, etc. in the same sentence. If your AI agent can’t follow that, it fails at the exact moment users behave naturally.

Sarvam Vision (Document intelligence / OCR)

Sarvam also pushed into document understanding with Sarvam Vision — aiming beyond basic OCR to interpret real documents: multi-column layouts, tables, stamps, forms, and mathematical expressions. Reports in February 2026 cited strong scores like 84.3% on olmOCR-Bench (English subset) and around 93.28% on OmniDocBench v1.5 (English subset), emphasizing performance on complex layouts and formulas. (Coverage example: The Hans India.)

Source: moneycontrol.com

This chart illustrates Sarvam Vision’s benchmark performance, highlighting its accuracy in document intelligence compared to other leading models.

Why this matters: India runs on documents — IDs, certificates, land records, invoices, permits. If AI can extract and validate fields reliably (and in Indian scripts), it can reduce manual work, errors, and waiting times.

Bulbul V3 (Text-to-Speech for voice agents)

On the output side, Sarvam released Bulbul V3, a text-to-speech model aimed at natural, consistent voices for Indian languages — a core building block for voice agents. Reporting in early February 2026 mentioned 35+ voices and support for 11+ Indian languages, with plans to extend to all 22 scheduled languages. (Coverage example: Indian Express; API page: Sarvam TTS API.)

From Demos to Deployment: State-Level Partnerships

The most interesting signal is not a benchmark — it’s deployment intent. In 2026, multiple state-level announcements connected Sarvam with “sovereign compute” projects: building data centers and AI facilities intended for government and public-service use cases.

Odisha (50MW AI-optimized facility)

On February 6, 2026, Odisha signed an MoU with Sarvam AI tied to a 50MW AI-optimized facility, framed as a sovereign AI hub and a boost for high-skilled jobs. (Coverage example: Economic Times.)

Tamil Nadu (₹10,000 crore “Sovereign AI Park”)

On January 13, 2026, Tamil Nadu announced an agreement with Sarvam AI for a ₹10,000 crore sovereign AI park initiative, often described as a “full-stack” AI ecosystem. The messaging emphasized that data, models and compute remain within the state’s trust boundary — a classic “sovereign AI” framing. (Coverage example: Economic Times, plus The New Indian Express.)

Important nuance: building AI infrastructure is hard. Announcements are one thing — sustained operations, procurement, governance, and security are another. The real test will be whether these initiatives deliver reliable public services while minimizing bias, leakage of sensitive data, and vendor lock-in.

What to Watch Next

• Production evidence: real deployments (helplines, document workflows, call center automation) with measurable accuracy and latency.
• Language coverage: expansion from 11+ supported languages to all 22 scheduled languages across speech and TTS.
• Safety & bias: how models handle sensitive queries across India’s social and cultural complexity.
• Developer accessibility: stable APIs, pricing, and documentation to compete with global platforms.

Conclusion

Sarvam AI’s story is bigger than “an Indian ChatGPT.” It’s about building the plumbing for AI in India: speech recognition that works in code-mixed conversations, document intelligence that handles real paperwork, and voices that sound natural in local languages. If sovereign AI becomes a reality, it will likely be powered by these less glamorous building blocks — deployed quietly in services millions depend on.

Frequently Asked Questions about Sarvam AI