Indian AI Startup Open-Sources New Models
Indian AI startup Sarvam AI has released two new open-source multilingual reasoning models. The models, built on government supercomputers, support 22 Indian languages and are designed to be lightweight enough to run on a laptop. This move positions Sarvam to compete globally in coding and agentic AI tasks.
The startup was founded in August 2023 by Vivek Raghavan, who played a key role in building India's digital public infrastructure like Aadhaar, and Pratyush Kumar, a former Microsoft and IBM researcher known for his work in open-source Indic AI at AI4Bharat. Their stated goal is to build AI for "Bharat," focusing on the vast non-English speaking population of India. Sarvam AI secured a landmark $41 million Series A funding round in December 2023, the largest for an Indian AI startup at that stage. The round was led by Lightspeed Venture Partners, with participation from Peak XV Partners (formerly Sequoia Capital India) and noted Silicon Valley investor Khosla Ventures, an early backer of OpenAI. The models, named Sarvam 30B and Sarvam 105B, were developed as part of the government's ₹10,372-crore IndiaAI Mission. This initiative provided Sarvam with access to government-supported GPU compute infrastructure, with technical support from Nvidia and data center operator Yotta, showcasing a significant public-private partnership. Both models utilize a Mixture-of-Experts (MoE) architecture, which activates only a fraction of the models' parameters at a time to enhance efficiency and lower compute costs. The larger 105B model features a 128,000-token context window, positioning it for complex reasoning and agentic workflows, while the 30B model is optimized for real-time conversational AI. Benchmark tests indicate the 105B model is competitive with similarly sized global models like Qwen3-Next and pt-oss 120B. It notably outperforms models such as DeepSeek R1 and Gemini 2.5 Flash on the Tau 2 Bench for agentic reasoning and task completion. These open-source models are not just research projects; they already power Sarvam's commercial products. The 105B model is the foundation for "Indus," an AI assistant for complex tasks, while the 30B model runs "Samvaad," the company's conversational agent platform. The models are available for download on Hugging Face and AIKosh under the Apache 2.0 license.
