India’s ‘Sovereign AI’ Goes Global as Sarvam Releases Open-Weight Models

The spotlights in New Delhi’s Bharat Mandapam convention center were fixed on Prime Minister Narendra Modi as he slipped on a pair of AI‑powered smart glasses in mid‑February, flashing a quick smile to the cameras.

On a nearby stage at the India AI Impact Summit 2026, executives from Bengaluru‑based startup Sarvam AI were making a different kind of pitch: two large language models they described as built in India, for India, and central to the country’s push for “sovereign AI.”

“Design and develop in India. Deliver to the world. Deliver to humanity,” Modi told delegates at the summit’s leaders’ session the next day.

Less than three weeks later, anyone with a decent internet connection could download the same “sovereign” models from open repositories and reuse them under one of the most permissive software licenses in the world.

The releases—Sarvam‑30B and Sarvam‑105B—along with a new Indian‑language chatbot called Indus, mark the most visible test yet of India’s bet that state‑backed, open‑weight artificial intelligence can deliver both national control and global influence.

From summit showcase to open release

Sarvam announced the two models on Feb. 18 at the New Delhi summit, which the government billed as a flagship event for the Global South. The startup described Sarvam‑30B and Sarvam‑105B as large language models trained from scratch on data heavy in Indian languages and local context, optimized for reasoning, coding and long documents.

Two days later, on Feb. 20, Sarvam rolled out Indus, a consumer chatbot app that runs on the 105‑billion‑parameter model. The app, available on the web and on Apple and Google app stores, initially opened in limited beta with a wait list. Indian tech outlets quickly dubbed it a “desi rival” to OpenAI’s ChatGPT and Google’s Gemini.

Indus supports conversational text and voice in English and 22 Indian languages, including Hindi, Tamil and Bengali, and can switch languages mid‑conversation. In leaked system instructions posted by developers online, the service identifies itself as “Indus, an AI assistant made by Sarvam AI, running on the Sarvam 105B model.”

The consequential move: open weights under Apache‑2.0

The more consequential move came on March 6. That day, Sarvam published the full model weights for Sarvam‑30B and Sarvam‑105B on popular hosting platforms such as Hugging Face and Indian service AIKosh under the Apache‑2.0 license.

In practical terms, that license allows anyone—inside or outside India—to download, modify and commercially deploy the models, provided they respect attribution and a handful of other conditions.

Sarvam describes both systems as mixture‑of‑experts models, a sparse architecture in which only a subset of parameters is active for each token. The headline 105‑billion‑parameter model uses about 10.3 billion active parameters per token and supports a 128,000‑token context window, allowing it to process long documents and multi‑step reasoning chains. The 30‑billion‑parameter model is tuned for efficiency, with about 2.4 billion active parameters and long‑context handling.

The company says the models were trained entirely in India using domestic compute and proprietary datasets with heavy coverage of Indian languages, code‑mixed text such as Hinglish, and local knowledge. In technical documentation, Sarvam reports that Sarvam‑105B scores above 90 on the MMLU benchmark, a widely cited test of general knowledge and reasoning, and posts strong results on coding and math benchmarks.

In Indian media interviews, executives have claimed the 105B model outperforms Chinese model DeepSeek R1 and Google’s Gemini 2.5 Flash on certain Indian‑language technical tasks and said it scored 70 out of 75 on this year’s Joint Entrance Examination (JEE) Mains exam in a single attempt, and 75 out of 75 across two attempts, in company‑run tests.

Those claims have not yet been independently verified. As of early March, the models did not appear on some of the most watched open leaderboards maintained by third‑party platforms, a gap that analysts say makes it difficult to compare them directly with other leading systems.

“India can train a sovereign model but still cannot prove it works,” one Forbes columnist wrote, calling for more transparent, third‑party evaluations of the models’ capabilities, particularly in Indian languages.

Government backing and the IndiaAI Mission

The government has provided ample backing for Sarvam’s efforts. The startup, founded in August 2023 by former Aadhaar architect Vivek Raghavan and AI researcher Pratyush Kumar, was selected in April 2025 as one of the companies to build India’s national language model under the IndiaAI Mission.

That mission, approved by the Cabinet in March 2024 with a budget outlay of 10,371.92 crore rupees (about $1.25 billion), aims to build indigenous AI compute infrastructure, curated data platforms, startup support and a dedicated AI safety institute under the Ministry of Electronics and Information Technology.

Officials have said Sarvam was granted access to about 4,000 high‑end graphics processing units and more than 200 crore rupees’ worth of compute credits to train its frontier‑scale models. State governments have followed with their own partnerships. Tamil Nadu has announced a 10,000‑crore‑rupee “Digital Sangam” sovereign AI park with Sarvam and the Indian Institute of Technology Madras, including a 20‑megawatt AI‑optimized data center. Odisha has signed a memorandum of understanding around a sovereign AI capacity hub that forms part of a $2.3 billion package of AI and infrastructure commitments.

Union Electronics and IT Minister Ashwini Vaishnaw has repeatedly presented Sarvam as a key beneficiary of this push.

“We are confident that Sarvam’s models will be competitive with global models,” he said in an interview last year, framing the IndiaAI Mission as an effort to build a “frugal, sovereign and scalable” ecosystem.

A sovereignty paradox

The government’s use of the term “sovereign AI,” however, sits uneasily with Sarvam’s decision to release its flagship models under Apache‑2.0. Policy documents and summit materials have described sovereign models as those trained on domestic compute, using domestic data, and controlled by Indian entities for use in public services and critical sectors.

In technical terms, once weights are released under a permissive license, downstream control is limited. Developers around the world have already converted Sarvam’s models into alternative formats suitable for local deployment and integrated them into open‑source inference engines. Community members have also published “uncensored” variants of the 30B model that strip out safety and content‑filtering layers, illustrating how alignment is often implemented at the system‑prompt or fine‑tuning level rather than in the underlying weights.

AI researchers and policymakers say that openness is a double‑edged sword.

The case for openness

On one hand, it allows Indian startups, universities and small firms to build applications without relying on expensive proprietary APIs from U.S. technology companies. Local developers can fine‑tune the models for education, agriculture, health or translation, in languages and dialects that mainstream systems often handle poorly.

India has hundreds of millions of internet users more comfortable in non‑English languages, and many lack the literacy or confidence to navigate English‑only interfaces. Sarvam and its backers argue that models trained from scratch on Indian‑language data, combined with voice interfaces, can help close that gap.

The risks of open weights

On the other hand, open weights mean capable models can be downloaded, modified and hosted entirely beyond India’s jurisdiction. Security experts have warned that large, high‑quality Indian‑language models could be used to automate influence campaigns, targeted harassment or fraud in local languages, and to generate synthetic audio and video that mimic political figures.

India’s AI safety institute is still being set up, and publicly available documentation offers little detail on how open‑weight models like Sarvam’s will be governed once they are in the wild. The company’s own content policies for Indus include restrictions on hate speech, misinformation and illegal activity, but those controls do not automatically transfer to third‑party deployments.

Scrutiny from the developer community

Sarvam has also faced questions within the open‑source community. Last year, an online discussion accused the company of artificially inflating download statistics for an earlier model hosted on Hugging Face. The allegation has not been formally adjudicated but has contributed to skepticism among some developers, who say they want clearer, independently verifiable metrics on the new systems.

What comes next

For now, India’s wager is that the benefits outweigh the risks. By funding domestic compute capacity and putting competitive open‑weight models into circulation, officials hope to seed a local industry that can serve Indian users and export to other developing countries wary of depending solely on U.S. or Chinese AI providers.

The coming months will test that strategy. Indus will need to prove itself against entrenched global chatbots on the smartphones of Indian consumers. Researchers will scrutinize Sarvam‑30B and Sarvam‑105B on public leaderboards and in real‑world deployments. Regulators will confront how to manage powerful, freely downloadable models in dozens of languages.

Back at Bharat Mandapam, the summit banners have long come down. But the tension on display there—between a political promise of national control and a technical choice to open the code to the world—is only beginning to play out.