Indian AI firm Sarvam has launched a new suite of large language models, emphasizing its commitment to open-source technology amidst a burgeoning competitive landscape dominated by U.S. and Chinese companies. Unveiled at the India AI Impact Summit in New Delhi, this initiative is part of India’s broader strategy to decrease dependency on foreign AI platforms and to customize solutions for local languages and applications.
The latest offerings include models with 30-billion and 105-billion parameters, along with advancements in text-to-speech, speech-to-text, and document analysis capabilities. These models represent a significant leap from the company’s earlier 2-billion-parameter Sarvam 1 model released in October 2024.
Key Features of Sarvam’s New Models
- Parameter Configuration: The 30B and 105B models employ a mixture-of-experts architecture, activating only a portion of their parameters at any given time, which dramatically reduces computing costs.
- Context Windows: The 30B model is optimized for real-time conversations with a context window of 32,000 tokens, while the 105B model features an extensive 128,000-token window, designed for complex reasoning tasks.
- Training Approach: Both models have been trained from scratch on vast datasets, with the 30B model utilizing approximately 16 trillion tokens and the 105B model leveraging trillions of tokens across various Indian languages.
Sarvam asserts that these models are crafted for real-time applications, including voice assistants and chat systems tailored for Indian languages, positioning them in direct competition with established models like OpenAI’s GPT-OSS-120B and Alibaba’s Qwen-3-Next-80B.
The training process was supported by the government-backed IndiaAI Mission, leveraging infrastructure from data center operator Yotta and technical expertise from Nvidia. Pratyush Kumar, co-founder of Sarvam, emphasized the company’s focus on practical applications over mere size, stating, “We want to understand the tasks that truly matter and build for them.”
Looking ahead, Sarvam plans to open-source its 30B and 105B models, although details on the availability of training data and comprehensive training code remain unclear. The company also aims to develop specialized AI systems, including coding-focused models and enterprise solutions under the Sarvam for Work initiative, as well as a conversational AI platform named Samvaad.
Founded in 2023, Sarvam has garnered over $50 million in funding, with notable investors including Lightspeed Venture Partners, Khosla Ventures, and Peak XV Partners (formerly Sequoia Capital India).
