Sarvam AI Unveils Two New Large Language Models

Bengaluru-based startup Sarvam AI unveiled two new large language models on Tuesday at the India AI Impact Summit in New Delhi, marking a significant step in India's drive to develop homegrown artificial intelligence systems tailored to local needs.

The models, a 30-billion-parameter version and a 105-billion-parameter version, were trained from scratch using a mixture-of-experts architecture. This design activates only a subset of parameters for each input, enhancing efficiency by reducing computational demands. The smaller model activates about 1 billion parameters per token and supports a 32,000-token context window, making it suitable for real-time conversational applications. It was pretrained on approximately 16 trillion tokens of text data. The larger model activates around 9 billion parameters and handles a 128,000-token context window, enabling more complex reasoning and multistep tasks.

Sarvam AI co-founder Pratyush Kumar, who presented the models during a keynote address, highlighted their performance on industry benchmarks. He stated that the 105-billion-parameter model outperforms DeepSeek's R1, a 600-billion-parameter reasoning system, while being more...