India’s Homegrown LLM Effort Gains Momentum: Meet the 4 Startups Leading the Charge
Backed by government push and GPU investments, Indian AI firms aim to deliver foundational models tailored for local needs
India’s First Indigenous LLMs Target Six-Month Timeline
India is set to unveil its first indigenously developed large language models (LLMs) within the next six to eight months, according to Union IT Minister Ashwini Vaishnaw, as of January 30, 2025.
- The initiative marks India’s strategic response to the global AI race, particularly China’s DeepSeek, an open-source LLM built at a fraction of global development costs.
- Indian officials insist this is part of a pre-planned national AI strategy, with the government having already accepted GPU bids far exceeding initial projections — from 10,000 to over 18,000 units.
“We have kept GPU bids open. If not for LLMs, where else could this infrastructure be used?” said a senior official.
Four Startups Selected to Build India’s AI Foundation
As of May 31, the government has shortlisted four AI startups to develop foundational LLMs, each with a unique focus:
1. Sarvam: Versatile, Scalable, On-Device LLM Suite
Founded by Vivek Raghavan and Pratyush Kumar in Bengaluru, Sarvam is building a three-tiered LLM family:
- Sarvam-Large for advanced reasoning and text generation.
- Sarvam-Small for real-time interactions.
- Sarvam-Edge for on-device tasks that demand compact model footprints.
This suite is aimed at enabling scalable AI deployment across cloud and edge applications.
2. Soket AI: Open-Source LLM for India’s Languages
Founded by Abhishek Upperwal, Soket AI has deep roots in AI and HPC, pivoting to NLP and foundational models in 2019.
- The Bengaluru-based firm is working on a 120-billion parameter open-source model, optimized for India’s linguistic diversity.
- Target sectors include defence, healthcare, and education.
Soket plans to begin with 1–2 billion parameter models, then scale progressively to avoid resource waste. The vision includes building towards India’s first AGI (Artificial General Intelligence) platform.
3. Gnani.ai: Voice-to-Voice LLM with Emotional Intelligence
Gnani.ai, also headquartered in Bengaluru, is tackling a voice-first approach to LLMs.
- The startup is developing real-time speech-to-speech AI models, starting with 14 billion parameters and expanding to 70 billion over a year.
- These models aim for low-latency autonomous conversations, with capabilities such as emotion-aware voice interactions.
“We want to deliver instant, emotionally intelligent responses,” said Ganesh Gopalan, co-founder and CEO.
4. Gan.ai: Super-Human Text-to-Speech Foundation Model
Founded by Stanford alumni Suvrat Bhooshan and Parth Sarthi, Gan.ai is developing a multi-lingual 70-billion parameter foundation model.
- Its goal: deliver text-to-speech quality beyond current global benchmarks.
- Although still bootstrapped, the startup is receiving support through the national AI mission and is yet to finalize an India base.
Government officials see Gan.ai’s innovations as potentially leapfrogging existing international models in natural speech synthesis.
Building India’s AI Stack: From Chips to Language Intelligence
This LLM initiative aligns with India’s ₹10,000 crore AI fund, which includes:
- 50,000 GPUs allocated for model training.
- A focus on building AI services that cater to Indian languages and use cases.
- Encouraging open-source, sovereign AI development.
The motivation, partly sparked by Sam Altman’s claim that ChatGPT would remain unmatched, has fueled India’s tech leaders to develop AI that solves real-world, local problems at scale.
