Names, versions and owners change fast. Checked October 2026.
In plain words
Sarvam AI is a company, not a single app. It was started in Bengaluru in 2023 to build AI for India, especially for the many people who would rather speak or type in Hindi, Tamil, Bengali, Marathi or another Indian language than in English.
It makes several kinds of model. Some are large language models that chat, write and reason. Others turn speech into text (speech recognition), turn text into natural speech (text-to-speech), or read scanned forms and documents. Together they can power things like a voice helpline that answers in the caller's own language.
In 2025 the Indian government chose Sarvam to build a home-grown foundation model under its IndiaAI Mission, and in 2026 the company released its larger models with open weights, trained in India, so others can download and build on them.
Everyday example
A translator at the bank counter
Picture a small-town bank where the forms are in English, but the customer speaks only Kannada. A good counter assistant listens in Kannada, reads the form, explains it back in Kannada and fills it in. Sarvam's models aim to do each of those jobs: hear speech, read the document, understand the question and answer aloud in the customer's language.
Where the example breaks: A human assistant knows the customer and the bank's rules. The models only predict likely words and sounds, so they can mishear, mistranslate or be confidently wrong.
Latestas of October 2026
- Main language models
- Sarvam 105B and 30B, open weights since March 2026; updated 105B versions announced at Sarvam Epoch in August 2026
- App
- Indus, opened in limited beta on 20 February 2026 as a way to try the 105B model, and since expanded into a workspace for AI agents
- Newest speech models
- Saaras V4 speech recognition (25 September 2026) and Bulbul V4 Flash text-to-speech (13 October 2026)
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
Building AI for Indian languages raises a few specific problems, and Sarvam's models are designed around them.
A tokenizer for Indian scripts
Language models read text as tokens. Models built mainly for English split Indian scripts into many tiny pieces, which makes them slower and costlier. Sarvam's tokenizer covers all 22 scheduled Indian languages across 12 scripts. See tokenization.
Trained from scratch in India
Its 30B and 105B models, named for their billions of parameters, were trained from scratch, rather than fine-tuned from a foreign model, on computing provided under the IndiaAI Mission.
Speech for real phone lines
Many users call rather than type. Sarvam's speech models are built for code-mixed speech, such as Hindi with English words, and for low-quality telephone audio.
Open weights
The 30B and 105B models were released in March 2026 under the permissive Apache 2.0 licence, on Hugging Face and on AIKosh, an Indian government AI platform.
- Company
- Sarvam AI, Bengaluru, founded in 2023 by Vivek Raghavan and Pratyush Kumar
- IndiaAI Mission
- Selected in April 2025 to build an indigenous foundation model
- Open models
- Sarvam 30B and Sarvam 105B, weights released 6 March 2026 under Apache 2.0
- Languages
- Built around the 22 scheduled Indian languages and English
Sarvam is one of several Indian groups building models for Indian languages under the IndiaAI Mission.
Questions people ask
Is Sarvam AI India's own ChatGPT?
Partly. Sarvam builds large language models and an app, Indus, where people can try them, much as people use ChatGPT. But it focuses on Indian languages and also makes speech and document models. It is a separate Indian company, not linked to OpenAI.
Who founded Sarvam AI?
Vivek Raghavan and Pratyush Kumar founded it in Bengaluru in 2023. Both had earlier worked with AI4Bharat, a group at IIT Madras that builds language technology for Indian languages.
Is Sarvam AI open source?
Some of it. Its 30B and 105B language models were released in March 2026 with open weights under the Apache 2.0 licence, so anyone can download and use them. Other models are offered through Sarvam's own apps and its API for businesses.
Which languages does Sarvam AI support?
Its work is built around English and the 22 scheduled languages of India, such as Hindi, Bengali, Tamil, Telugu, Marathi and Kannada. Support differs by model: some speech models cover all 22, others cover fewer.