The AI Impact Summit has been making headlines for many reasons. Be it Chinese robodogs being wrongly called our own, or tech leaders getting awkward on stage while holding hands. However, there were some
announcements that need our attention. For instance, India launched three made-in-India sovereign AI LLM models that were unveiled and proudly showcased to the world. These three models are built by India companies Sarvam AI, Gnani AI, and BharatGen.
The three models - Sarvam AI's twin LLMs, Gnani.ai's Vachana voice stack, and BharatGen's Param2, are products of the IndiaAI Mission, a government initiative launched in 2024 with a budget of over Rs. 10,000 crore. The idea was to offer AI tools to India's 1.4 billion people speaking 22 scheduled languages, carry smartphones into low-bandwidth villages, and interact with government services in ways that ChatGPT was never trained to understand.
Google CEO Sundar Pichai, who was in the audience, put it simply, "The developer energy I find in India is second to none. What you're talking about is actually happening. India is very well positioned."
Let's dive into each sovereign LLM to see what they offer:
1. Sarvam AI
Of the three launches, Sarvam AI's announcement was perhaps the most watched. The Bengaluru-based startup was handpicked by the Indian government in April 2025 from a shortlist of 67 companies to build the country's first indigenous foundational LLM under the IndiaAI Mission. At the Summit, it delivered on that promise.
Sarvam unveiled two models side by side - a 30-billion-parameter model named Vikram (after space pioneer Vikram Sarabhai) built for real-time conversations with a 32,000-token context window, and a 105-billion-parameter model built for heavier, more complex reasoning tasks with a 128,000-token window. Both were trained entirely in India, not fine-tuned adaptations of American or Chinese models, but built from scratch on sovereign compute, using domestic infrastructure, and packed with Indian language data.
Benchmarks shown at the launch placed the models competitively against Gemma 27B, Mistral-32-24B, Nemotron-30B, Qwen-30B, and GPT-OSS-20B on tasks spanning mathematical reasoning, coding accuracy, and general problem-solving. Sarvam's co-founder Vivek Raghavan captured the philosophy of the project, "Sovereignty matters much more in AI than building the biggest models."
The models come in three variants being developed under the broader sovereign LLM roadmap. Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge, perhaps the most striking, which runs entirely on consumer devices without internet connectivity. For a country where connectivity still drops off a cliff outside urban centres, an AI that works offline in Hindi, Kannada, Telugu, or Gujarati is not a luxury. It's infrastructure.
How to access it: Sarvam's models are available via API at sarvam.ai, with a developer playground for testing. The sovereign LLM stack is also being deployed through government partnerships, including with UIDAI (Aadhaar) for voice-based citizen services.
2. Gnani.ai
If Sarvam is building India's brain, Gnani.ai is building its voice. The Bengaluru-based conversational AI company - already processing 10 million calls a day before this Summit - arrived with a system that can clone a human voice in 12 Indian languages using fewer than 10 seconds of audio.
The product is called Vachana TTS (Text-to-Speech), and it is the kind of technology that reframes what 'accessible AI' actually means. Feed it a short audio sample - a government official's recorded greeting, a teacher's voice, a healthcare worker's introduction - and Vachana clones that voice's pitch, timbre, and speaking style, then deploys it across every language it supports without the voice losing its identity. A district collector can deliver emergency alerts in Hindi, Tamil, and Odia, all in their own voice, without recording each one separately.
The supported languages span the breadth of India - Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Marathi, Punjabi, Odia, Assamese, and Indian English. The model achieves a Mean Opinion Score of 4.23 and a character error rate below 0.6 percent, numbers that place it firmly in human-quality territory. Clients including the Tata Group and Air India are already deploying it.
Alongside TTS, Gnani.ai also launched Vachana STT (Speech-to-Text), trained on over one million hours of real-world audio across more than 1,000 domains. It handles the messy audio realities of India, noisy call centres, compressed telephony, regional accents, and code-mixed speech (the everyday blending of English with local languages) - without needing domain-specific fine-tuning. Word error rates sit below 5 percent for Hindi and Indian English, and below 10 percent for most other supported languages.
The use cases are wide - public announcements, emergency alerts, educational content delivery, AI-powered call centres, and voice-first access to government schemes for citizens with low literacy or visual impairments. Gnani estimates the technology can reduce call-centre costs by 40 percent to 60 percent.
How to access it: Vachana TTS is available immediately via API. On-premises deployment is available for regulated sectors such as banking and government. Visit gnani.ai for developer access.
3. BharatGen
The third entrant is the most unusual. BharatGen is led by IIT Bombay and backed by the IndiaAI Mission to the tune of Rs. 900 crore - making it the largest single beneficiary of government AI funding to date. Its mandate is to build AI that is not just multilingual, but legally and culturally coherent with India.
At the Summit, BharatGen launched Param2 17B MoE - a 17-billion-parameter multilingual foundational model built on a Mixture of Experts (MoE) architecture, trained across all 22 scheduled Indian languages, and developed in collaboration with NVIDIA. The MoE design is significant. Rather than using the full model for every query, it routes each input to the most relevant 'expert' sub-network, making it more efficient and more accurate for specialised domains. The result is a model that understands not just the words of Indian languages, but their legal frameworks, governance contexts, and cultural registers.
Param2's target domains are governance, defence, agriculture, healthcare, and education. These are not sectors that ChatGPT was designed to serve well in India, they require nuance about how a village panchayat works, how crop insurance schemes are structured, or how healthcare schemes operate in a specific state. BharatGen's academic backing means it has been built with those specifics in mind, rather than as an afterthought.
How to access it: BharatGen is releasing Param2 and its documentation via Hugging Face, making it openly available for developers to fine-tune and build on. Researchers and enterprises can access the model weights directly for India-centric applications.
