Indian AI startup Sarvam has unveiled a new generation of large language models, marking a bold strategic move to prove that smaller, efficient open-source AI systems can compete with larger, more expensive models developed by U.S. and Chinese tech giants. The announcement came at the India AI Impact Summit in New Delhi, aligning with India’s national push to reduce dependence on foreign AI platforms and build technology tailored to local languages and real-world needs. The new lineup includes a 30-billion-parameter model and a 105-billion-parameter model, alongside specialized systems such as a text-to-speech model, a speech-to-text model, and a vision model designed to interpret documents. This represents a significant leap from Sarvam’s earlier 2-billion-parameter Sarvam 1 model, launched in October 2024. Both new models use a mixture-of-experts architecture, which activates only a subset of parameters during inference, dramatically lowering computational costs. The 30B model features a 32,000-token context window, optimized for real-time conversations, while the 105B model supports a 128,000-token window, enabling advanced, multi-step reasoning tasks. Sarvam emphasized that these models were trained from scratch, not fine-tuned from existing open-source systems. The 30B model was pre-trained on approximately 16 trillion tokens of text, while the 105B model was trained on trillions of tokens spanning multiple Indian languages, enhancing its ability to serve diverse linguistic communities. Designed for practical deployment, the models aim to power voice-based assistants, chat systems, and other real-time applications in Indian languages. The 105B model is positioned to compete with OpenAI’s GPT-OSS-120B and Alibaba’s Qwen-3-Next-80B, while the 30B model is benchmarked against Google’s Gemma 27B and OpenAI’s GPT-OSS-20B. The development was made possible through India’s government-backed IndiaAI Mission, which provided computing infrastructure via data center operator Yotta and technical support from Nvidia. Sarvam’s leadership stressed a thoughtful, application-driven approach to scaling, rejecting the idea of chasing raw model size for its own sake. “We want to be mindful in how we do the scaling,” said co-founder Pratyush Kumar during the launch. “We don’t want to do the scaling mindlessly. We want to understand the tasks which really matter at scale and go and build for them.” Sarvam plans to open-source both the 30B and 105B models, though it has not yet confirmed whether the training data or full training code will be publicly available. Beyond the core models, the company is developing specialized AI systems, including coding-focused models and enterprise tools under a new product line called Sarvam for Work, as well as a conversational AI agent platform named Samvaad. Founded in 2023, Sarvam has raised over $50 million in funding from prominent investors including Lightspeed Venture Partners, Khosla Ventures, and Peak XV Partners (formerly Sequoia Capital India). The company’s latest move underscores a growing belief in India’s potential to become a global hub for practical, localized, and efficient AI innovation.