Indian AI lab Sarvam’s new models are a major bet on the viability ofopen-source AI

Indian AI lab Sarvam launched new large language models at the India AI Impact Summit, aiming to compete with U.S. and Chinese firms by providing open-source options tailored to local languages. Their 30B and 105B parameter models utilize a mixture-of-experts architecture and were trained on extensive data sources. Sarvam's approach focuses on efficiency and real-world applications with plans to open-source their models.

Key Points

• Sarvam unveiled new language models with 30 billion and 105 billion parameters.
• Models aim to capture market share from larger competitors by providing efficient open-source alternatives.
• New models trained from scratch on trillions of tokens focusing on Indian languages.
• The 30B model offers real-time conversational support, while the 105B model supports complex reasoning.
• Models were developed with government support under the IndiaAI Mission, and infrastructure from Yotta and Nvidia.
• Sarvam plans to open-source both models and focus on practical applications rather than sheer scale.

Relevance

• The launch of Sarvam aligns with current trends emphasizing open-source solutions and language localization in AI, crucial for developing economies.
• As of 2025, there is an increasing trend towards AI models that are accessible and tailored for specific markets, reducing dependence on foreign technologies.
• Historically, India has aimed to boost its technological independence, making this development part of a larger national strategy.

Sarvam's new AI models represent a strategic move in the growing open-source AI landscape, focusing on efficiency and localization, which could have significant implications for the competitive dynamics of AI technologies in India and beyond.