Sarvam AI unveils two new LLMs; says 105B model surpasses DeepSeek's R1 and Google's Gemini Flash on key benchmarks

Bengaluru-based startup Sarvam AI has introduced two new large language models, as it expands its role in India’s effort to build domestic AI systems.

Speaking at the India AI Impact Summit in Delhi, Sarvam AI Co-founder Pratyush Kumar said the company has trained a 30-billion-parameter model and a 105-billion-parameter model from scratch, using a mixture-of-experts (MoE) architecture to balance scale and efficiency.

The MoE architecture boosts efficiency by activating only a subset of the total parameters for each input, rather than the whole model.

Kumar said Sarvam had previously developed a 3-billion-parameter dense model but it had to scale further. The 30B model, he explained, activates only 1 billion parameters per token despite having 30 billion parameters in total. He said this reduces inference costs and improves efficiency, particularly for reasoning tasks. The model supports a 32,000-token context window and was trained on 16 trillion tokens.

Efficiency, Kumar said, is central to the company’s approach as it aims to build AI that delivers population-scale impact.

Sarvam also presented a 105-billion-parameter MoE model that activates 9 billion parameters and supports a 128,000-token context window. Kumar said the system is designed for more complex reasoning and agentic use cases.

Kumar said, on most benchmarks, the 105B model outperforms DeepSeek’s R1, a 600-billion-parameter reasoning model released last year. It is cheaper than Google’s Gemini Flash and surpasses it on several benchmarks, he added. On Indian language tasks, he said, the model performs better than Gemini 2.5 Flash.

The launches come as the government-backed IndiaAI Mission pushes for sovereign foundational models.

The mission, supported by a Rs 10,000-crore fund, has allocated Rs 111 crore for GPU subsidies so far.

Sarvam has received nearly Rs 99 crore in subsidies and secured 4,096 NVIDIA H100 SXM GPUs through Yotta Data Services, making it the largest beneficiary of the IndianAI Mission to date. The company was earlier chosen as the first startup under the mission to build India’s foundational AI model.

Sarvam was founded in July 2023 by Vivek Raghavan and Kumar, who had previously worked at AI4Bharat, the research lab of IIT Madras. The company offers research-led model development and enterprise deployment tools.

Beyond models, Sarvam is also expanding into hardware.

Kumar said Sarvam plans to launch 'Sarvam Kaze', a smart eyewear product designed and built in India, by May. He said the device would allow developers to build applications on top of it and is part of the company’s push to build sovereignty across technology layers, including user-facing devices.

At the summit, Sarvam also announced partnerships with HMD to bring AI capabilities to feature phones; with Qualcomm to deploy generative AI solutions across smartphones, PCs, wearables, XR, IoT, automotive, and data centres; and with Bosch to integrate AI into car panels.