Sarvam Launches Indus Chat App as Interface for its AI Models

AI start-up Sarvam on Friday announced the launch of its Indus chat app for web and mobile users.

Indus, currently available in limited beta, is an interface that allows users to interact with Sarvam’s AI models. It is powered by Sarvam’s 105-billion parameter sovereign model. However, the interface is still smaller in scale compared to the advanced models used in global consumer chat applications.

Sarvam stated at this stage, it is deliberately focusing on improving accuracy, usefulness, efficiency, and alignment with Indian needs before developing larger foundational models.

AI in India

The launch comes at a time when global AI companies such as OpenAI, Anthropic, Google (Gemini), and X are competing to expand their user base in India.

Recently, OpenAI CEO Sam Altman said that ChatGPT has more than 100 million weekly active users in India. Anthropic also said India accounts for 5.8% of total usage of its Claude chatbot, making it its second-largest market after the United States.

Both OpenAI and Anthropic have announced plans to open offices in India and partner with major Indian companies such as Infosys and Tata Group.

The company said, “True AI sovereignty means building the entire technology stack from the ground up. The first step was to create foundational models in India. Indus is the next step, showing that India can also own the data and interface layers.”

It added that Indus is an early release and is being rolled out gradually due to limited computing capacity. Some users may initially be placed on a waitlist, but access will be expanded over time.

Sarvam also emphasised that sovereign AI must be a collective national effort. The company said it should be built with the participation of developers, researchers, local experts, creators, artists, and everyday users who best understand India’s needs.

Sarvam 30B & 105B

Recently, Sarvam announced the launch of two new large language models, a 30-billion-parameter model and a 105-billion-parameter model.

The company unveiled the models at the India AI Impact Summit in Delhi, stating that both use a mixture-of-experts (MoE) architecture designed to improve efficiency while maintaining strong performance across reasoning, programming, and tool-use tasks.

The 30B model activates only 1 billion parameters per token, despite having 30 billion parameters in total.

“It is actually a mixture-of-experts model. We have a 30-billion-parameter model, but in generating every output token, it activates only 1 billion parameters,” said Pratyush Kumar, cofounder of Sarvam.

He added, “If you look at the thinking budget, Sarvam 30B significantly outperforms others at both the 8K and 16K scales compared to the latest models released at the same size.”