New Delhi, Feb 18 (NationPress) The AI startup Sarvam AI, based in Bengaluru, announced the launch of two new large language models as part of India's initiative to establish its own sovereign artificial intelligence capabilities.

This revelation occurred during the India AI Impact Summit, where the company detailed that both models were developed from the ground up utilizing a mixture-of-experts (MoE) architecture to enhance efficiency and performance.

The first model, named Sarvam 30B, comprises 30 billion parameters, but only 1 billion parameters are activated for each output token it generates.

Co-founder Pratyush Kumar emphasized that this MoE design minimizes inference costs while boosting efficiency, particularly for reasoning and complex tasks.

He noted that the 30B model excels in thinking and reasoning benchmarks at both 8K and 16K scales when benchmarked against other models of comparable size.

The Sarvam 30B model features a 32,000-token context window and has been trained on a staggering 16 trillion tokens.

Kumar reiterated that efficiency is at the forefront of the company's vision, aiming to make AI accessible on a large scale throughout India.

Additionally, the company launched a more advanced model, the 105-billion-parameter model, which is tailored for intricate reasoning and agent-based tasks.

This model utilizes 9 billion parameters and accommodates a 128,000-token context window, enabling it to process more complex instructions and sustain longer conversations.

Kumar drew comparisons between the new 105B model and leading global systems, asserting that it surpasses DeepSeek’s DeepSeek R1, which had 600 billion parameters at its launch last year, across various benchmarks.

He also mentioned that this model is more cost-effective than Google’s Gemini Flash, while outperforming it in numerous benchmarks.

According to him, when compared to Gemini 2.5 Flash, Sarvam's model displays superior performance on tasks involving Indian languages.

This launch coincides with India's intensified efforts to create foundational AI models that cater to multilingual and large-scale public applications.

Backed by the government, the IndiaAI Mission is supported by a significant Rs 10,000 crore fund, aiming to decrease reliance on foreign AI systems and foster domestic innovation.

To date, the mission has allocated Rs 111 crore in GPU subsidies, with Sarvam AI emerging as the primary beneficiary, acquiring 4,096 NVIDIA H100 SXM GPUs through Yotta Data Services and receiving nearly Rs 99 crore in subsidies.

The startup was previously selected as the inaugural company to develop India’s foundational AI model under this mission.

Sarvam AI was established in July 2023 by Vivek Raghavan and Pratyush Kumar, both of whom were involved with AI4Bharat, an initiative supported by Infosys co-founder Nandan Nilekani.