Sarvam AI Opens Samvaad to Developers After 350M Real Conversations

Sarvam AI, India's full-stack sovereign AI company, has opened its voice agent platform to the public. What was previously locked behind enterprise waitlists and high-volume contracts is now available to any developer, startup, or SMB looking to build a voice-powered AI agent. The platform, called Sarvam Samvaad, has already quietly powered more than 350 million conversations in production, making this less of a debut and more of a graduation.

From enterprise gatekeeper to self-serve

Until now, access to Sarvam's conversational AI agents was largely limited to enterprises and customers with high conversation volumes. The public launch introduces a self-serve model with free credits and usage-based pricing, letting startups, SMBs, developers, and individuals build and deploy localized voice agents without a sales conversation. Free-tier users face certain usage caps, with paid plans expected for those who want to run without limits.

Enterprise AI procurement cycles can stretch for months. A self-serve funnel compresses that to minutes, which is the point.

What the platform actually does

Samvaad lets enterprises and government agencies deploy voice and text agents that operate fluently across 11 Indian languages. Its key technical claims: sub-500ms latency for real-time voice with contact-centre-grade audio quality, multi-agent orchestration for complex workflows, and cross-channel memory spanning voice calls, WhatsApp, and web.

The stack is genuinely end-to-end. Here is what it bundles together:

• Omnichannel reach: Interactions across telephone, WhatsApp, web, and apps.
• Persistent memory: Agents remember past conversations and context across sessions and channels.
• Enterprise integrations: Direct connectors to CRM, core banking, and payment systems, so agents can handle appointment booking, payment follow-ups, cart recovery, collections, and inbound support without a human in the loop.
• Multilingual fluency: Sarvam 30B powers the platform. Translation runs in two modes: Sarvam-Translate for formal 22-language coverage, and Mayura for colloquial and code-mixed Hinglish.
• Deployment flexibility: Cloud, VPC, or on-premises.

The code-switching capability deserves a closer look. Generic TTS systems handle mixed-language speech by detecting language boundaries and routing segments to separate engines, producing audible pauses and accent shifts. Bulbul V3 handles code-switching at the model level, generating the entire mixed-language sentence in one pass with no seams at the language boundary. For a country where "Aapka order dispatch ho gaya hai, expected delivery by tomorrow" is normal speech, that matters enormously.

The numbers that earned this moment

The 350 million conversations figure comes from real enterprise deployments, not demo traffic. Sarvam's voice agents collected data from seventeen million farmers as part of an exercise commissioned by India's Ministry of Agriculture. A separate voice campaign reached forty-five million insurance policyholders. These are population-scale deployments.

The moat is Indic depth: roughly a 90% win rate on Indian-language benchmarks, native-script OCR, and code-mixed speech, rather than English frontier parity. OpenAI and Google have Indian language support; Sarvam has Indian language expertise, and the deliberate choice shows in the benchmarks.

Who built this, and why now

Sarvam AI is a Bengaluru-based sovereign AI company founded by Vivek Raghavan and Pratyush Kumar in August 2023 to build large language models, speech systems, and AI infrastructure for India's 22 official languages and 1.4 billion people. Both founders came out of AI4Bharat, a research collective with a decade of Indic language dataset work behind it.

The timing of the public launch tracks with the fundraise. Sarvam entered the unicorn club after raising $234 million in the first close of its Series B, reaching a $1.5 billion valuation. $150 million of that came from HCLTech as lead strategic investor, with Bessemer Venture Partners joining alongside existing backers Khosla Ventures and Peak XV Partners. With that capital in hand, opening the platform to public developers is the logical way to build the ecosystem that justifies the valuation.

Sarvam is the first major Indian generative-AI company whose lead investor is an Indian conglomerate rather than a Silicon Valley fund. That is a structural signal as much as a financial one. Running a sovereign LLM stack is closer to a geopolitical infrastructure decision than a technology one, similar to how India runs its own payment network (UPI) rather than operating purely on Visa and Mastercard.

An asymmetric fight with ElevenLabs

By opening the platform, Sarvam is positioning itself against global voice AI providers like ElevenLabs while leaning on its multilingual, India-focused stack. The move puts it in direct competition with ElevenAgents, launched in September 2025. ElevenLabs currently runs multiple pricing tiers globally, from free plans with limited call minutes to premium enterprise-grade agent plans.

The competitive dynamics here are asymmetric. ElevenLabs is a global product with multilingual support bolted on. Sarvam is an India-first product with global ambitions. Its MoE models (an architecture where specialized sub-networks handle different inputs, keeping inference efficient) are trained from scratch on domestic compute and cover all 22 Indian languages with tokenization far more efficient than generic multilingual models. For Indian enterprises and government bodies, that depth is hard to replicate from the outside.

Domestic competition is also real. Krutrim, founded by Ola co-founder Bhavish Aggarwal, was India's first AI unicorn and is building its own sovereign language model stack and GPU infrastructure. For now, the two companies are targeting different layers of the stack.

What comes next

The fresh capital is earmarked for next-generation models focused on agentic AI, coding, and cybersecurity applications. Between 30% and 50% of the funds will go toward GPU procurement, according to Raghavan. The second close of the Series B, targeting a total of $300 million, is still pending.

For builders, the practical question is whether the self-serve platform delivers on its latency and reliability promises once public traffic hits it. Sarvam's infrastructure currently serves 10 billion+ tokens with median latency under 100ms and a 99.9% uptime SLA. Those are strong baseline numbers. Holding them as the user base expands from a curated enterprise list to open public access will be the real test.

For anyone building fintech, healthtech, agritech, or government-facing products for Indian users, Sarvam's APIs for speech-to-text, text-to-speech, translation, and Indian-language LLM are a credible foundation. The platform is open, the 350 million conversations are the proof of concept, and the interesting question is what the developer community builds on top of it.