What is Sarvam AI? The models, the languages and how developers use them
Sarvam AI is a Bengaluru-based Indian AI company that says it has built a full-stack AI platform developed, deployed and governed entirely in India. Developers use its APIs for Indian-language speech and text: Bulbul v3 for text to speech, Saaras v3 for speech to text, Sarvam-105B for chat, and Mayura and Sarvam-Translate for translation. You sign up, create an API key and pay per use in rupees.
Prices as on 17 September 2026. Figures were checked against the provider's official pages on that date. Providers change prices and rules without notice, so confirm on the official page before you budget. Convo360 adds no markup: you pay the provider directly with your own account. Convo360 is not affiliated with Sarvam AI. Sarvam, Bulbul, Saaras, Mayura and other product names are trademarks of their owner. Details were read from Sarvam's official website and docs on 17 September 2026 and may change.
- Sarvam says it was founded in August 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar, and builds AI models for Indian languages.
- The current model names are Bulbul v3 (voice), Saaras v3 (transcription), Sarvam-105B (chat), Mayura and Sarvam-Translate (translation) and Sarvam Vision (documents).
- Bulbul v3 and Sarvam-105B cover 11 languages; Saaras v3 and Sarvam-Translate cover 23, according to Sarvam's docs.
- Saarika v2.5, Sarvam-M and Sarvam-30B are marked deprecated in Sarvam's docs as on 17 September 2026.
- In Convo360 you add your own Sarvam key under Settings → Provider Keys and choose a Sarvam voice on the agent.
Who Sarvam AI is, in its own words
This page was checked against Sarvam's own website and developer docs as on 17 September 2026. Convo360 is not affiliated with Sarvam AI, so every descriptive claim below is attributed to Sarvam.
Sarvam's about page carries the line "AI for all from India". It says the company was founded by Dr. Vivek Raghavan and Dr. Pratyush Kumar in August 2023, and describes what it has built as a full-stack AI platform, with everything developed, deployed and governed entirely in India. The website footer lists a Bengaluru address.
Sarvam offers two kinds of things:
- Products for enterprises, which its site lists as Voice Agents, Content Agents, Doc Agents, Work Agents, Coding Agents and Sarvam Edge.
- Model APIs for developers: text to speech, speech to text, translation, chat completion and document digitisation, all behind a single API key.
If you are building a WhatsApp bot or an AI phone agent for Indian customers, the Model APIs are the part that matters. The rest of this page sticks to them.
The Sarvam model family and what each model does
Sarvam's model names carry version numbers, and older versions get retired, so it pays to know the current ones. These are the models listed in Sarvam's models documentation.
| Model | Job | What Sarvam says about it | Status |
|---|---|---|---|
| Bulbul v3 | Text to speech | 30+ speaker voices, pace control, up to 2,500 characters per request | Active |
| Saaras v3 | Speech to text | Five output modes: transcribe, translate, verbatim, translit and codemix | Recommended |
| Sarvam-105B | Chat LLM | Flagship 105B parameter model, with a conversations variant for real-time dialogue and voice agents | Active |
| Mayura | Translation | Translation with context preservation | Active |
| Sarvam-Translate | Translation | Extended translation across more languages | Active |
| Sarvam Vision | Document intelligence | OCR and structured output from PDFs and images | Active |
A few version notes will save you confusion when you read older tutorials:
- The Bulbul model ID is bulbul:v3. Sarvam's docs label bulbul:v2 as legacy.
- The Saaras page also lists saaras:v4 as the latest model, which adds Global English. Saaras v3 remains the default and recommended model.
- Saarika v2.5, the older speech to text model, is marked deprecated, with advice to migrate to Saaras v3.
- Sarvam-M and Sarvam-30B, the older chat models, are marked deprecated, with advice to migrate to Sarvam-105B.
- Sarvam also serves a few open-source models in beta. Its docs state these are not tuned for Indian languages.
Which Indian languages Sarvam supports
Language coverage differs by model, so check the model you plan to use and not only the headline number.
| Set | Models | Languages |
|---|---|---|
| 11 languages | Bulbul v3, Mayura, Sarvam-105B | Hindi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Odia and English |
| 23 languages | Saaras v3, Sarvam-Translate, Sarvam Vision | The 11 above plus Assamese, Urdu, Nepali, Konkani, Kashmiri, Sindhi, Sanskrit, Santali, Manipuri, Bodo, Maithili and Dogri |
In practice this means a caller can be transcribed in more languages than the agent can speak back in. If your customers speak Assamese or Konkani, plan the spoken reply carefully and test before launch.
Sarvam's docs say Saaras is built for code-mixed speech and 8 kHz telephony audio, which is what real Indian phone calls sound like. For Bulbul, the docs warn that romanised Indic text reduces output quality. Write "आपका order confirm हो गया है" and not "Aapka order confirm ho gaya hai". More on language choice in AI voice for Tamil, Telugu and Marathi.
How developers get access
Access is self-serve. According to Sarvam's authentication page:
- Sign up on the Sarvam dashboard.
- Open the API Keys section and generate a key manually. The value is shown only once, so save it immediately.
- Send the key in the api-subscription-key header. A standard Bearer token header is also accepted, which helps with OpenAI-compatible tools.
Sarvam publishes official Python and JavaScript SDKs, plus REST, WebSocket streaming and batch endpoints. Its docs also list integration guides for LiveKit, Pipecat, Vapi, Twilio and Exotel.
Two limits are worth knowing early. Organisation-level keys are not supported yet, so credits sit on a single account. Rate limits also apply to the account as a whole and rise with the plan tier.
Billing is prepaid and in rupees. Sarvam gives ₹100 of free credits on signup and says credits never expire. Bulbul v3 is listed at ₹30 per 10,000 characters and speech to text at ₹30 per hour of audio. The full table, the plan tiers and a cost-per-call formula are in our Sarvam AI pricing guide.
Where Sarvam fits in an Indian-language voice or chat agent
A voice agent is a loop of three jobs: hear the caller, decide what to say, and say it. Sarvam's models documentation lays out the same pipeline for a multilingual voice assistant, with one of its models for each job:
- Hear: Saaras v3 turns the caller's Hindi, Tamil or Hinglish speech into text.
- Decide: Sarvam-105B, or any other LLM, writes the reply.
- Speak: Bulbul v3 turns that reply into speech in one of its voices.
You do not have to take all three from one company. Each job is a separate API, so you can pair one provider's LLM with another provider's voice. For a chat agent on WhatsApp, the speech models drop out and only the LLM and translation matter.
Sarvam also lists its own Voice Agents product for building and deploying conversational agents. That is a separate route from using the Model APIs inside another platform. Which one suits you depends on how much of the CRM, flows and channels you want in one place.
How Sarvam plugs into Convo360
Convo360 is an India-first platform for AI voice calls and WhatsApp chatbots, and it is BYOK. Sarvam is one of four voice providers you can choose per agent, alongside Smallest.ai, ElevenLabs and OpenAI. The agent's brain runs on your own OpenAI, Anthropic Claude or Google Gemini key.
- Generate an API key in your Sarvam dashboard.
- In Convo360, go to Settings → Provider Keys and paste it. Keys are encrypted at rest.
- Open the voice AI agent, select Sarvam as the voice provider and choose a voice. Convo360 fetches the voice list dynamically, so new Sarvam voices appear without any typing of IDs.
- Make a test call, then attach the agent to your Exotel, Plivo, Twilio or SIP number.
Sarvam bills your account directly, and Convo360 adds zero markup. If the key stops working or credits run out, Provider Health monitoring shows a clear alert naming the provider, so you are not left guessing why calls went silent.
Voice calls are on the Pro plan at ₹2,999 per month, excluding 18% GST, with a 14-day free trial. Setup steps are in the voice docs, and a side-by-side view of providers is in Indian voices for AI calls.
Example conversation
Try a Sarvam voice on your AI agent, free for 14 days
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
Is Sarvam AI an Indian company?
Yes. Sarvam's about page says it was founded in August 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar, and that its platform is developed, deployed and governed entirely in India. Its website lists a Bengaluru address.
What is Bulbul v3?
Bulbul v3 is Sarvam's current text to speech model. Sarvam's docs describe 30+ speaker voices, pace control from 0.5x to 2.0x and support for 11 languages, which are ten Indian languages plus English. The older bulbul:v2 is labelled legacy.
What is the difference between Saaras and Saarika?
Both are Sarvam speech to text models. Saarika v2.5 is the older one, covering 11 languages, and Sarvam's docs mark it deprecated. Saaras v3 is the recommended model, covering 23 languages with transcribe, translate, verbatim, translit and codemix modes.
Is Sarvam-M still available?
Sarvam's models page, as read on 17 September 2026, marks Sarvam-M and Sarvam-30B as deprecated and says Sarvam-M is no longer available through the API. Sarvam advises migrating to Sarvam-105B, its current flagship chat model.
Is Sarvam AI free to use?
Not fully. Sarvam's docs say every new user gets ₹100 worth of free credits on signup, usable on any API, and that credits never expire. After that you buy credits and pay per character, per second of audio or per token.
How do I use Sarvam AI with Convo360?
Create an API key in your Sarvam dashboard, paste it in Convo360 under Settings → Provider Keys, then select Sarvam as the voice provider on your voice agent and pick a voice. Sarvam bills you directly and Convo360 adds zero markup.
Official sources
- Sarvam AI: About us (official)
- Sarvam AI docs: Models overview (official)
- Sarvam AI docs: Bulbul text to speech model (official)
- Sarvam AI docs: Saaras speech to text model (official)
- Sarvam AI docs: Authentication and API keys (official)
- Sarvam AI docs: Pricing (official)
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.