Sarvam AI in 2026: who owns it, its models and the Indian languages it supports

Sarvam AI builds language, speech, translation and document models for Indian languages, and it is one of twelve organisations the Indian government is supporting to build sovereign foundation models.¹ That mix of a private company and a national programme is why people search for whether Sarvam is a government company. This profile answers that from the primary sources, then covers its funding, its current models, the languages they support, how developers get access, and how it compares with global models on the benchmarks that are public.

It is written as of 4 November 2026. Sarvam ships quickly, so check its models page before relying on a version number here. How assistants find and cite sources in general is covered in the AI search guide, and running a site in several languages in the international SEO guide.

Sarvam was founded in Bengaluru in 2023 by Vivek Raghavan and Pratyush Kumar.²³ Raghavan spent about a decade building Indian digital public infrastructure, including Aadhaar. Kumar came from academic and corporate research and is a founder of AI4Bharat, an open-source research initiative on Indian-language AI. The name is Sanskrit for "all".³

The company trades as Sarvam AI. Its terms of service name the legal entity as Axonwise Private Limited, with its registered office in Bengaluru and Indian law governing the contract.⁴ If you are checking company records, that is the name to search for.

Sarvam is a private limited company owned by its founders and investors. The government is involved in three ways: it funds Sarvam's model training through the IndiaAI Mission, it may take a small stake that has been reported but not confirmed, and several government bodies are Sarvam customers or partners.

The IndiaAI Mission selection

On 26 April 2025 the Government of India, under the IndiaAI Mission, selected Sarvam to build India's sovereign large language model, with AI4Bharat at IIT Madras as a collaborator.⁵ A month later the Ministry of Electronics and IT described the plan as an open-source model of 120 billion parameters for governance and public services.⁶ By February 2026 the government put its financial and compute support for Sarvam at ₹246.72 crore, as one of 12 organisations selected under the mission's Innovation Centre pillar.⁷ Information submitted to the Lok Sabha in March 2026 said the assistance is provided in a calibrated manner to cover compute usage costs.¹ Sarvam says its 30B and 105B models were trained entirely in India on compute provided under the mission.⁸

The reported government stake

The mission's call for proposals, published in January 2025, offered grants and compute credits paid against milestones, and said additional funding could come as equity-based investment through mutual agreement.⁹ In June 2026 The Economic Times reported that the Centre was considering a 1 to 2% stake in Sarvam, arising from the compute support and likely to be acquired through convertible instruments. Business Standard still described it as a reported proposal in July 2026.² I have not found a statement from MeitY or Sarvam confirming it. If it goes ahead, a stake of that size would make the government a small minority shareholder in a company that remains private.

Government bodies as customers and partners

UIDAI, which runs Aadhaar, is working with Sarvam on voice services, fraud alerts and support in 10 Indian languages, on UIDAI's own on-premise infrastructure. The Odisha government is setting up a 50MW AI compute hub with Sarvam, and Tamil Nadu and IIT Madras are developing an AI research park anchored by a 20MW data centre with it.⁷ In October 2026 Sarvam signed a memorandum of understanding with Headquarters Integrated Defence Staff, an organisation under the Ministry of Defence, setting up a framework for AI projects across the Army, Navy and Air Force.¹⁰ Its Series B announcement also cites voice agents that collected data from 17 million farmers for the Ministry of Agriculture and Farmers Welfare.¹¹

HCLTech's $150 million buys a stake of about 10.5%.¹³ Sarvam says the Series B money will pay for research on its next frontier model for agentic, coding and cybersecurity work, and for compute.¹¹

Sarvam's earlier models were small or adapted from other labs' models. Sarvam 1, released in October 2024, had 2 billion parameters.¹⁴ Sarvam-M, a 24B model, was built on Mistral Small and post-trained on Indian languages,¹⁵ and is no longer available through the API.¹⁶ The current line-up centres on two language models trained from scratch, unveiled at the India AI Impact Summit in February 2026 and released as open weights in March.¹⁴⁸

At its Epoch conference in Bengaluru on 30 July 2026, Sarvam said it is building a model of more than a trillion parameters from scratch in India, expected to go live within six months.²⁵ Its models page listed no such model on 4 November 2026.²⁴

Indus is Sarvam's own app. It launched in February 2026 as a limited-beta interface to the 105B model.²⁶ At Epoch, Sarvam brought its voice, work, coding, content and document agents and its inference service under the Indus platform with one credit system,²⁵ and developers manage their API keys in the Indus dashboard.²⁷

Sarvam's documentation says its speech, text and document models are trained for 22 Indian languages and English.²⁷ The 22 are India's scheduled languages:⁷ Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu and Urdu.²⁰

Coverage differs by product, which matters more than the headline number. Speech recognition and Sarvam Translate cover all 22.²⁰²² The flagship chat model is tuned for the 10 most-spoken languages and English, and text to speech and short-text translation cover 10 Indian languages and English: Hindi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi and Odia.¹⁶²¹²³ So the current line-up can transcribe a caller speaking Konkani or Maithili, but its text-to-speech models do not cover those languages.

Sarvam also builds for the way Indians type. Its Indian-language benchmark tests every prompt in native script and in romanised Latin script, which Sarvam treats as the colloquial form common in messaging and online communication,⁸ and the 105B model accepts native, romanised and code-mixed input.¹⁶

• API and SDKs. REST, WebSocket and batch endpoints cover transcription, speech synthesis, translation, chat and document extraction, with official Python and JavaScript SDKs.²⁷ The chat API is compatible with OpenAI's chat completions format.¹⁶ Sarvam says more than a million developers have registered on its platform.²⁵
• Pricing. New accounts get ₹100 of free credits, and credits do not expire. On pay as you go, speech to text costs ₹0.49 a minute, translation ₹2 per 1,000 characters, and Sarvam 105B ₹15 per million input tokens and ₹60 per million output tokens. Monthly plans from ₹9,999 to ₹1,99,999 lower the unit rates.²⁸
• Open weights. Sarvam 105B, Sarvam 30B and Prism are on Hugging Face under Apache 2.0, and Sarvam Translate under GPL-3.0.^17181922 The 30B and 105B are also on AIKosh, the government platform that makes IndiaAI models available to Indian startups and researchers.⁸¹
• Hosting. Sarvam's API serves its own models alongside open-weight models from other labs, including Gemma 4, GLM and DeepSeek, from data centres in India.²⁸²⁹ GLM and DeepSeek come from Chinese labs, which the guide to Chinese AI models covers. Its work agents can also be deployed in a private cloud, on premises or fully air-gapped.³⁰

Sarvam's own large models were trained in India. Sarvam 30B and 105B used IndiaAI Mission compute, with Nvidia contributing software libraries,⁸ and a review of Indian foundation models identifies that compute as Yotta's Shakti H100 cluster.³¹ Sarvam's inference platform runs from data centres in India on what it calls the first fleet of Blackwell GPUs in the country.²⁹

The documented Google connection is Gemma, Google's family of open models. Sarvam Translate is a fine-tune of Gemma 3 4B,²² and in July 2025 Google said it was collaborating closely with Sarvam, which had built a long-form translation model using Gemma.³² Sarvam also serves Gemma 4 on its own API.²⁸ I could not find a published agreement under which Sarvam runs its models on Google Cloud.

Microsoft announced a collaboration in February 2024 under which Sarvam would use Azure AI and Azure infrastructure, including Azure OpenAI Service and Azure Machine Learning, to build and host its Indic voice model, starting in Hindi. The two companies also said they would research better support for Indian languages in Microsoft's own models.³³ That announcement predates the sovereign models. HCLTech, the lead Series B investor, says it will bring Sarvam's models together with its own global presence to create a full-stack AI platform for enterprises and governments.¹¹

Most comparisons come from Sarvam, and independent evidence is thin. A review of publicly reported benchmark results by two authors at UIDAI, published in August 2026, found that most Indian-language results are reported by the developers themselves, that different models report different benchmark suites, and that no single Indic benchmark has been adopted across Indian developers.³¹ UIDAI is a Sarvam partner, though the paper only collates published scores.

On Sarvam's own pairwise test, its models beat open models of similar size in Indian languages, though on MILU, also reported by Sarvam, Qwen3-30B-Thinking scored higher than Sarvam 30B. Its call-data benchmark favours Sarvam 105B over GPT 5.4 Mini and Gemini 3.5 Flash. The independent Indian-language test I found that included Sarvam 105B found a smaller general model, Gemma 4 31B, more reliable, though that test covered only four languages. The government said in March 2026 that the sovereign models launched at the summit had in some cases beaten leading frontier models on specific tasks,¹ without naming the tasks. If you are choosing a model for an Indian-language product, test the shortlist on your own content in your own languages.

The audience is large. The IAMAI and Kantar Internet in India 2024 report, based on more than 90,000 samples across 35 states and union territories, found that 870 million internet users, 98% of the total, accessed the internet in Indic languages, and that 57% of urban users preferred content in vernacular languages.³⁶ Search is moving with them. Google's AI Mode launched in India in English and Hindi, and in October 2025 Google began adding Bengali, Kannada, Malayalam, Marathi, Tamil, Telugu and Urdu.³⁷

Sarvam's tools are most useful where Indian languages are spoken or processed at volume: customer calls, transcription, document digitisation and translation. Its voice agents are priced from ₹3 a minute.²⁸ For search, an Indian-language market needs the same work as any other language market, plus a few details specific to India.

Translating pages with Sarvam

Google's spam policies do not ban machine translation. They do list scraping content to generate many pages, including through automated translation, where little value is provided, as an example of scaled content abuse.³⁸ A Sarvam translation of a service page, corrected by a fluent speaker and adapted to the market, is a normal way to localise. Running a whole blog through the API into 22 languages and publishing it unread is the pattern the policy describes. Localising content for international SEO sets out a review process by page tier.

Language codes and hreflang for Indian languages

Do not copy Sarvam's language codes into hreflang. Google's hreflang supports only ISO 639-1 language codes, with optional ISO 3166-1 region codes.³⁹ Sarvam's API writes Odia as od-IN,²⁰ but Odia's ISO 639-1 code is or,⁴⁰ so the hreflang value is or-IN. Six of the 22 scheduled languages, Bodo, Dogri, Konkani, Maithili, Manipuri and Santali, have no ISO 639-1 code at all,⁴⁰ so by Google's documentation they cannot be targeted with hreflang.

For pages in those six languages, the visible text does the work. Google determines a page's language from its visible content and does not use code-level signals such as the lang attribute.⁴¹ The lang attribute should still carry the right code, which for these languages is a three-letter one: the W3C notes that the language subtag registry uses a two-letter code where one exists and a three-letter code otherwise.⁴² How to audit hreflang covers checking the rest of the set.

Romanised and code-mixed searches

Romanised Indian-language text is common enough in messaging and online writing that Sarvam tests every prompt in it.⁸ Keyword research for a Hindi or Tamil market should include romanised and mixed-language versions of the main queries alongside native script, and the pages may need to answer both. Which form people use varies by topic and audience, so check the search data for each market before deciding.

For visibility in AI answers, track the assistants your customers already use, in each language you serve, the way measuring visibility in AI search describes.