[Exclusive] Sarvam AI Could Launch its Indigenous AI Model on Republic Day 2026

The Government of India, under its IndiaAI Mission, selected Bengaluru-based Sarvam AI to build the country’s sovereign large language model (LLM).

The model will be developed with reasoning capabilities, voice support, and fluency in Indian languages, using local infrastructure and dedicated government compute-resources — including over 4,000 Nvidia H100 GPUs — for large-scale deployment.

As a part of this arrangement, a government body ( the Digital India Corporation) has acquired an equity stake in Sarvam AI in return. Now, sources familiar with the matter have told The Left Shift that a few government officials have held discussions to release the model on Republic Day 2026.

Previously, IT Minister Union Minister Ashwini Vaishnaw has stated that Sarvam AI will launch the model early next year.

So far, there have been no official communications from either the government or Sarvam AI about a release date for the model. Previously, Sarvam AI co-founder Vivek Raghavan had told The Left Shift that the model could be announced by the end of the year.

Sarvam AI also faces multiple challenges in building India’s foundational AI model under the IndiaAI Mission — including limited access to high-quality multilingual data, delay in access to compute infrastructure, and the complexity of scaling for diverse Indian languages, dialects, and accents while meeting global performance benchmarks and maintaining true model sovereignty. This could be the result of the delayed launch.

So far, the Bengaluru-based startup has launched a few open-source models; however, it has received widespread criticism from various sections of the Indian tech community. Earlier this year, it launched Sarvam M, a 24-billion-parameter open-weights hybrid language model built on top of Mistral Small. 

Critics note its reliance on Mistral Small rather than a fully home-grown architecture, and they argue that its improvements are marginal in many benchmarks, especially in English.

Early user uptake was also low: just a few hundred downloads on Hugging Face in the first days sparked comparisons to models from other countries that saw much higher popularity.

Some observers say the model lacks standout features and doesn’t solve “hard problems,” seeing it more as a stepping stone than a breakthrough.

The Indian government has selected three more AI startups—Soket AI, Gnani.ai, and Gan.ai—in the first batch to develop large-scale foundational models as part of its ambitious IndiaAI Mission.

On September 12, 2025,  the government selected BharatGen, Tech Mahindra, and Fractal, along with Avataar.ai, ZenteiQ.ai, Genloop, NeuroDX (IntelliHealth), and Shodh AI, for the second phase of its foundation model initiative.