Imagine asking your phone a question in Hindi, Marathi, Tamil or Bengali — and getting an answer back in your own language, from an AI built in India, for India. That is the promise CoRover has been building towards since 2016, and this week it took a big step onto the national stage: three of its BharatGPT models are now listed on AIKosh, the Government of India’s national platform for AI datasets and models under the IndiaAI Mission.
CoRover, a Gurugram-based conversational AI company, says the listing followed a review, testing and approval process by IndiaAI and the Ministry of Electronics and Information Technology (MeitY). It puts the company’s homegrown language and speech technology on the platform meant to power India’s own AI development — available to developers, enterprises and government teams building the next generation of Indian applications.
Three models, one mission: AI that speaks India
The first of the three is BharatGPT Mini, a 538-million-parameter foundational language model that the company says it trained from scratch. Small by the standards of today’s giant models, it is built to be nimble: it runs on-premises, on cloud, at the edge, or in hybrid setups — and even offline on low-end devices, without an internet connection. That matters enormously in a country where millions of people still come online through patchy networks and budget phones. The model supports 14 Indian languages, and it was unveiled at VivaTech in Paris in June 2025 by Union Minister Jitin Prasada.
The second, BharatGPT Instruct, is built on a Mixture-of-Experts architecture for text generation, coding, reasoning and enterprise tasks across Indic and global languages — the workhorse for companies that want to build assistants, automate support and put conversational AI into their products.
The third is ASR Hindi Conformer, which turns spoken Hindi into text. The company cites a Word Error Rate of 24.72% and a Character Error Rate of 21.78% on the AI4Bharat/Lahaja Hindi test set, results it describes as competitive with much larger speech recognition systems.
A voice assistant that understands India’s languages, wherever you are. (Illustration)
“The availability of our BharatGPT models on AIKosh marks an important step in making Indian-developed AI capabilities accessible through a national platform,” said Ankush Sabharwal, Founder and CEO of CoRover.ai.
From chatbots to sovereign AI
CoRover began life building enterprise chatbots, and its assistants now serve organisations including IRCTC, LIC and NPCI. BharatGPT, launched as India’s first indigenous large language model, grew out of that work — with voice support in more than 12 Indian languages and text in 22, built in collaboration with the National Hub of Language Technology under MeitY’s National Language Translation Mission.
The AIKosh listing is not the company’s only government milestone this year. CoRover says it is also building a Sovereign Agentic AI Framework for the Government of India, to support secure, compliant deployment of AI agents across departments and citizen services — and it recently won a MeitY contract to build a unified AI assistant platform for government, with DigiLocker first in line. Google DeepMind has also invested in the company, backing its push for sovereign Indian AI.
For CoRover, the vision has always been summed up in a phrase: make AI in India, and make AI work for India. With BharatGPT Mini, BharatGPT Instruct and ASR Hindi Conformer now sitting on the country’s own AI platform, that vision is one step closer to every developer and every citizen who wants to build — or simply talk — in their own language.
Keep reading: No Human Should Ever Climb Down There Again: Inside Project SHUDH, the Hyderabad Robots Taking Over India’s Sewers · The Heat Beneath the AI Boom: Inside KühlTherm, the Ahmedabad Startup Cooling India’s Data Centres With Liquid · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.







