Sarvam's Multilingual Stack: Vision OCR Beats Gemini 3 Pro

What Sarvam shipped in seven days

Between 5 and 11 February 2026, ahead of the IndiaAI Impact Summit, Bengaluru-based Sarvam AI ran a launch streak almost designed to intensify pressure on the rest of the Indic-AI field. Four production-grade releases in a single week, each one targeting a layer of the stack that, until recently, builders shipping multilingual products in India and the United Kingdom either bought from US labs at premium prices or stitched together from open-weight components that did not really speak Marathi, Tamil or Bengali. The headline numbers — Vision OCR at 84.3% on olmOCR-Bench, Bulbul V3 with thirty-five-plus voices across eleven languages, Sarvam Audio supporting automatic speech recognition for twenty-two Indian languages — are the easy part to memorise. The harder, and for builders more important, question is what you can now do that you could not last quarter.

This piece focuses on the product stack and the migration story. If you want the funding angle — Sarvam's $350M Series C at a $1.5B valuation, the NVIDIA / Accel / HCLTech investor list, the IndiaAI Mission ₹10,370 crore programme that anointed Sarvam as the first start-up to receive sovereign-AI support — that is a separate read. Here we are looking at olmOCR-Bench numbers, voice-AI use cases, ASR coverage gaps, and the practical question of whether an Indian agritech or a UK fintech serving diaspora customers should actually rebuild on the Sarvam stack today, hold their nose with OpenAI and Google for another quarter, or split-route between the two.

Launch	What it does	Headline metric	Builder relevance
Vision OCR	Document parsing, including Indic scripts and complex forms	84.3% on olmOCR-Bench	KYC, e-government, regulated financial document review
Bulbul V3	Text-to-speech across 11 Indian languages, 35+ voices	11 languages, 35+ voices	Voice-first apps, customer support, accessibility tooling
Sarvam Audio (ASR)	Automatic speech recognition for 22 Indian languages	22 Indian languages covered	Call-centre transcripts, voice notes, field-data capture
Stack integration	Pipeline binding OCR + ASR + TTS into single agent flows	One auth surface, one billing line	Multimodal agents serving multilingual users end-to-end

Vision OCR vs Gemini 3 Pro: where the 4-point gap actually shows up

The benchmark headline is genuinely striking. Sarvam's Vision OCR scored 84.3% on Allen AI's olmOCR-Bench, comfortably ahead of Google's Gemini 3 Pro at 80.2% and well ahead of ChatGPT at 69.8%. olmOCR-Bench is not a token-prediction toy benchmark — it is a deliberately nasty mixture of multi-column academic PDFs, scanned forms, equations, tables that bleed across pages, rotated photos of receipts, and noisy historical documents. Doing well on it correlates strongly with doing well on real document-understanding work.

Model	olmOCR-Bench score	Strength	Weakness vs Sarvam
Sarvam Vision OCR	84.3%	Indic scripts, regulated forms, mixed-language pages	Smaller open ecosystem (newer release)
Gemini 3 Pro	80.2%	Clean English layouts, multimodal grounding	Weaker on Devanagari + Tamil-Brahmi-derived scripts
ChatGPT (vision model)	69.8%	General reasoning over extracted text	Fails harder on rotated, low-DPI scans

Where the four-point lead actually shows up: Indic-script accuracy, where Sarvam's training emphasis pays off; tabular extraction on Indian government forms, where row-and-column inference under stamped overlays is the killer test; and mixed-language pages where a single document switches between English boilerplate and a regional-language signature block. Where the lead narrows or disappears: clean, single-language English academic PDFs, where Gemini 3 Pro's broader pre-training corpus matters more, and pages where the bottleneck is reasoning about extracted content rather than extraction itself.

Bulbul V3 voice: 35 voices, 11 languages, builder use-cases

Bulbul V3 is the speech-synthesis half of the new stack. The headline is thirty-five-plus voices across eleven Indian languages, but the more interesting fact for builders is the voice-banding: Sarvam ships distinct voices tuned for narration, conversational customer-service, and high-energy advertising registers, rather than a single neutral voice per language. That distinction matters when you are building, say, a Tamil interactive voice response system that has to feel local rather than synthetic.

Language	Voice variants	Best-fit use case
Hindi	~6	Mass-market customer support, e-learning
Tamil	~4	South-India fintech IVR, accessibility readers
Telugu	~4	Agritech advisory voice notes
Kannada	~3	Bengaluru civic-tech, healthcare triage
Malayalam	~3	Diaspora remittance UX, GCC outreach
Bengali	~3	Cross-border content (IN + Bangladesh)
Marathi	~3	Maharashtra public services, FMCG voice ads
Gujarati	~3	SME logistics, gold-loan voice flows
Punjabi	~3	UK / Canadian diaspora services
Odia	~2	State-government accessibility
Assamese	~2	North-east financial inclusion

The practical implication: a builder no longer has to choose between an English-centric voice provider with one underwhelming Hindi voice and a small local provider with patchy uptime. Bulbul V3 covers the breadth, and the voice variety means your product can pick a register that matches its tone — a reassuring older-man voice for a pension product, a younger, energetic voice for an edtech app for school-leavers — rather than shipping the same TTS personality to every audience.

ASR for 22 Indian languages: who needs it, who's shipping it

Sarvam Audio, the automatic speech recognition layer, supports twenty-two Indian languages. The number itself matters, but the more important fact is that this appears to be the first widely available commercial ASR system to ship parity coverage across the eight Constitution-listed major South Indian and North-East languages on the same SLA as Hindi. For comparison, Whisper-large covers roughly half of these with substantially higher word-error rates on accented or code-mixed speech, and most of the major proprietary US providers either skip the smaller languages or charge a premium for them.

The use-cases that move first are predictable: Indian call-centre transcription, where firms have been bolting Whisper to expensive correction layers; field-research voice-note capture for NGOs and agritech extension services; UPI-voice and conversational-banking products that need to handle code-mixed Hindi-English fluently; and accessibility tooling for visually impaired users in regional languages. Less obvious but important — UK-based diaspora services (council helplines, NHS interpreting workflows, remittance KYC voice-verification) where the ability to transcribe Tamil or Bengali on the same pipeline as English is a small operational unlock that meaningfully reduces interpreter cost.

Why this matters for IN + UK builders

Consider two builders who could plug into the same Sarvam stack tomorrow and ship a meaningfully better product than they could last quarter. First, an Indian agritech building voice-first interfaces for eleven-language farmers — historically one of the hardest segments to serve, because the user base cannot read instructions and your TTS budget per call has to be sub-rupee for the unit economics to work. Bulbul V3 plus Sarvam Audio gives this builder a single-vendor path: ASR captures the farmer's voice query in Marathi or Telugu, an LLM layer reasons over it, and Bulbul V3 voices the response back in the same language with a register that does not sound like a robot.

Now consider a UK or Commonwealth-facing fintech serving South Asian diaspora customers — Punjabi-speaking gold-loan customers in Birmingham, Tamil-speaking remittance senders in Croydon, Bengali-speaking NHS patients in Tower Hamlets. The same stack, with the same auth surface and the same billing line, lets this UK-headquartered builder ship a voice-first KYC or grievance flow in any of those languages without spinning up bespoke partnerships with eleven different regional vendors. The dual-market story is not a nice-to-have — it is the differentiator that lets a product team in either Bengaluru or London cover both customer pools without doubling their engineering surface area.

Open multimodal stack vs proprietary alternatives

The cost picture is qualitative rather than line-item — Sarvam's published price-per-1k-pages OCR is competitive with Google Document AI on English and meaningfully cheaper on Indic-script pages, where the proprietary providers either upcharge or fail outright. Voice synthesis comes in below ElevenLabs on per-character cost for the Indian languages, although ElevenLabs still wins on English voice expressiveness. The direction of travel — open-weight models, India-hosted inference, no per-seat licensing — looks structurally similar to the April open-weight wave with Llama 4, Mistral Small 4 and GLM-5.1, where the absolute cost is lower and the governance story is materially different.

The governance angle deserves its own paragraph. The IndiaAI Mission's ₹10,000 crore commitment (potentially doubling to ₹20,000 crore) and Sarvam's anointment as the first sovereign-AI start-up under the programme make this stack the politically obvious choice for Indian government and PSU procurement over the next eighteen months. Whether that is good news depends on which side of the table you sit on — for builders selling into government, it is a tailwind; for builders who simply want the best available model regardless of geography, it is a reminder that "best" is increasingly a function of where your customer is incorporated.

What to build with the Sarvam stack right now

The migration logic for an existing product team is short but not trivial. Step one — pick one of your existing flows where Indic-language quality is the bottleneck and the OpenAI / Google version is shipping noticeably worse outputs. Step two — wire up a parallel Sarvam endpoint for that single flow, using the same upstream prompts and the same eval harness, so you can compare side-by-side without rewriting your whole agent loop. Step three — instrument cost and quality both, in production, on real user traffic, for at least a fortnight; cherry-picked benchmark wins do not survive contact with messy users. Step four — make the migration call on data, not on press-release vibes.

Concrete project ideas that the new stack unlocks for builders not already shipping multilingual: a voice-first CRM for Indian SMEs that takes WhatsApp voice notes from field staff in any of the eleven Bulbul languages and produces structured CRM entries; a UK diaspora-facing legal-help tool that ingests scanned documents in Tamil or Bengali (Vision OCR) and produces a plain-English summary for a solicitor; a multilingual customer-support copilot that listens to inbound calls in twenty-two Indian languages (Sarvam Audio), suggests responses to a human agent, and reads them back to the customer in the matching language and register (Bulbul V3); an agritech extension app that delivers personalised crop advisories as voice notes in the farmer's first language. None of these were genuinely viable two quarters ago.

Three caveats before you migrate

1. Open-weight is not the same as open-source. Several Sarvam models are downloadable on Hugging Face under permissive terms, but full training data and complete training code are not always public. For most commercial uses that is fine — for licence-strict procurement, read each model card carefully and budget for legal review.
2. The benchmark wins do not generalise uniformly. Vision OCR's 4-point lead over Gemini 3 Pro is real on Indic-script and form-heavy pages, narrower on clean English layouts. Bulbul V3 is excellent on most South Asian languages but does not yet match ElevenLabs on English voice expressiveness. Pick the right tool for the right page.
3. Roadmap and stability risk is non-zero. Sarvam is now well-funded, but it is still a 2026-vintage start-up shipping at speed. Production-grade SLAs, multi-region failover, and long-term API-stability commitments are all in earlier maturity stages than what you get from Google or OpenAI. For mission-critical paths, design your client-side abstraction to allow fallback to a US lab. For everything else, the upside outweighs the risk.

The bigger picture is this: India now has a credible, end-to-end multimodal AI stack that takes Indic languages seriously, costs less than the US-lab alternative, and ships under a governance story that aligns with where Indian (and increasingly UK-procurement-rules-aware) customers want their data to live. The shift mirrors what is happening on the other side of the world — the UK's £500M sovereign AI fund is a different bet on the same thesis: that 2026's AI economy will not be a single global market but a federation of regional stacks tied to where compute, capital and customers live.

For Indian builders, this is the moment to stop treating Bengaluru AI as a niche choice and start treating it as the default for any flow that touches Indic languages. For UK builders, this is the moment to ask whether your South-Asian-language customer journeys are being served by a stack that was actually trained on those languages, or by a US-centric model that treats them as long-tail. And for the rest of the ecosystem — including the small but growing number of UK-based start-ups like Oolka shipping Indian-market AI agents — Sarvam's seven-day streak changes the cost-and-capability frontier in a way that is worth at least one production-pilot evaluation this quarter.

For deeper reading, see BusinessToday's profile on the Sarvam moment, the Founderpin company profile, Rest of World on India's frugal-AI bet, the Sarvam AI Wikipedia entry, and the olmOCR-Bench dataset card for benchmark methodology. For deployment patterns, see our Llama 4 deployment guide as an analogous open-weight playbook.