BharatGen and the Portfolio Logic of India's Sovereign-Model Bet

When the IndiaAI Mission selected which foundation-model efforts to fund from over 500 proposals, it did not pick a single national champion. It picked twelve. The logic is portfolio construction: rather than concentrate capital on one bet about what India's sovereign model should be, the state is spreading allocations across reasoning, voice and vertical-specific systems, accepting that several will underperform so that the cohort as a whole covers the capabilities that matter. BharatGen sits at the centre of that portfolio, and it shows what the deepest-funded position looks like.

BharatGen: the anchor allocation

BharatGen, the IIT Bombay-led consortium, presented Param2 at the Bharat Innovates event in June 2026, a model built for reasoning, coding and tool-calling across 22 languages. It follows Param-1, a 2.9-billion-parameter bilingual model trained on 5 trillion tokens, and the published roadmap runs from the 2B class up to a trillion-parameter target.

What distinguishes BharatGen is that it is building a full stack, not just a text model. The system spans Shrutam2 for automatic speech recognition (ASR), Sooktam2 for text-to-speech (TTS), and Patram for vision, the components required to make AI usable by populations that interact by voice and image rather than typed English. Tool-calling, the ability of Param2 to invoke external functions and APIs, is the bridge from a chatbot to an agent that can act, which is where enterprise and government value concentrates.

The funding reflects the priority. BharatGen received ₹1,058.52 crore, the largest sovereign allocation in the IndiaAI cohort by a wide margin, roughly four times the next-largest grant, alongside a collaboration with IBM. That is the state placing its biggest single chip on a full-stack, academically anchored effort.

The portfolio: de-risking across capability

The rest of the cohort fills out the bet. Fractal received ₹137.91 crore for an approximately 70B reasoning model plus Vaidya 2.0, a healthcare-focused system. Soket AI took ₹177.08 crore for Project EKA, an open-source 120B model targeting defence, agriculture and legal use. The voice-first allocations are the most telling: Gnani.ai received ₹177.27 crore for a 14B voice-to-voice model and its Inya VoiceOS, while Gan.AI received ₹110.03 crore for a 70B multilingual TTS system.

The concentration on voice is not incidental, it is the differentiation thesis for India. In a country with low English literacy and high linguistic diversity, the interface that scales is spoken language, not text. A voice-to-voice model that takes Marathi speech in and returns Marathi speech out, without a brittle transcribe-translate-synthesise pipeline, is far more deployable across rural India than a marginally smarter English text model. Funding two distinct voice efforts (Gnani's voice-to-voice, Gan's TTS) alongside reasoning models (Fractal, Soket) and verticals (Vaidya 2.0 for health) is the portfolio doing its job: covering the capability surface rather than betting the outcome on one architecture.

For investors, the structure carries a clear signal. The state is absorbing the early-stage risk of foundation-model building across the cohort, de-risking the layer that pure venture capital found hardest to underwrite, as Krutrim's retreat from frontier work demonstrated. The companies that convert these grants into deployed, revenue-generating systems, particularly in voice and regulated verticals where a sovereign model enjoys data-residency and procurement advantages, are the ones to watch. The grants are the floor, not the return. What matters next is which of the twelve build something the market, and not only the government, will pay for.