Sarvam AI And BharatGen Unveil Their Latest AI Models At The India AI Impact Summit 2026
The Sarvam AI’s 30B and 105B AI models are indigenously trained in India, while BharatGen’s Param2 17B MoE is built on NVIDIA AI Enterprise.
Published : February 20, 2026 at 7:38 PM IST
Hyderabad: Sarvam AI and BharatGen, two Indian Artificial Intelligence (AI) companies, unveiled their latest Large Language Models (LLMs) at the India AI Impact Summit 2026.
The Bengaluru-based Sarvam AI introduced two advanced models—Sarvam-30B and Sarvam-105B—both completely trained in India. Selected under the IndiaAI Mission in 2025, the models are designed to support multilingual deployments across government, enterprises, and developers, strengthening India’s data sovereignty.
Meanwhile, BharatGen unveiled its Param2 17B MoE AI model, which was developed in collaboration with NVIDIA. It is positioned as India’s flagship sovereign AI initiative. The model is optimised for multiple Indic languages and designed to serve critical sectors including governance, education, healthcare, agriculture, and enterprise solutions.
Sarvam 105B model follows the same MoE design, activating 9B parameters per token to combine large-scale capability with efficient execution.
— Sarvam AI (@SarvamAI) February 19, 2026
With a 128K context window, it is built for more demanding tasks including complex reasoning, agentic task completion, tool use, coding,… pic.twitter.com/gyvkuBpWIb
Sarvam AI’s Models
Sarvam AI’s 30B model is made for real-time conversational use cases, offering reduced latency and cost efficiency. It employs a mixture-of-experts (MoE) architecture, pre-trained on 16 trillion tokens, and supports up to a 32,000-token context window. With one billion activated parameters, it performs competitively against global benchmarks such as OpenAI’s GPT-OSS-20B and Google’s Gemma 27B.
Meanwhile, the 105B model is built for complex reasoning and advanced corporate operations. It features nine billion active parameters and a 128,000-token context window, which enables the 105B model to excel in long-form tasks, coding, and Indian language processing. It outperforms several leading open-source and closed-source frontier models, making it suitable for software engineering applications such as code generation and bug fixes.
We were honoured to have our Lead PI Prof Ganesh Ramakrishnan (@ganramkr) interact with the Hon’ble Prime Minister of India @narendramodi at the #IndiaAIImpactExpo2026, and share #BharatGen’s progress in building #SovereignAI LLMs across 22 Indian languages.@PMOIndia @IndiaDST pic.twitter.com/gR9TZzvT1f
— BharatGen (@BharatGen_Com) February 17, 2026
BharatGen’s AI model
The BharatGen’s AI model, Param2 17B MoE, is built on NVIDIA AI Enterprise. This end-to-end training pipeline uses NVIDIA’s Base Command Manager (BCM), which integrates with Slurm and supports workflows utilising NVIDIA’s open libraries NeMo-RL to manage and scale AI training. It ensures that the AI model offers scalability and performance.
BharatGen mentioned that the company is open-sourcing its AI models and tools via Hugging Face, allowing developers and enterprises to build and deploy India-focused AI solutions.
At the India AI Impact Summit 2026, Pratyush Kumar, cofounder and CEO of Sarvam AI, said that the Sarvam 105B AI model outperforms larger and more expensive devices like Gemini 2.5 Flash in Indian language performance.
While Rishi Bal, CEO of BharatGen, said, “We are extremely delighted at unveiling PARAM 2 17B at the AI Impact Summit in Bharat Mandapam and also at the partnership with NVIDIA. As a truly sovereign AI initiative, the launch and unveiling of PARAM 2 17B marks a pivotal moment for India, and we at BharatGen are elated at contributing to the country’s AI progress.”
