Sarvam AI Launches Homegrown Indian Language LLMs
Indian AI startup Sarvam has released two homegrown large language models with a focus on multilingual support for Indian languages. The launch event demonstrated the models' capabilities in languages like Hindi and Punjabi. This move signals a strategic focus on solving India-specific AI problems related to linguistic diversity rather than simply adapting global models for the domestic market.
- The company was co-founded by Dr. Vivek Raghavan, who has deep experience building India's digital public infrastructure, including Aadhaar, and Dr. Pratyush Kumar, a former researcher at Microsoft and IBM who also co-founded AI4Bharat, an open-source initiative for Indian language AI. - Sarvam AI secured one of the largest Series A funding rounds for an Indian AI startup, raising $41 million in December 2023 from investors including Lightspeed, Peak XV Partners, and Khosla Ventures. - One of their initial open-source releases is the "OpenHathi" series, a Hindi LLM built on Meta's Llama2-7B architecture, which demonstrates performance on par with GPT-3.5 for Indic languages and is available on Hugging Face for developers to build upon. - The company's technical strategy focuses on solving core problems for Indian languages, such as poor token efficiency, where Western models often require 4 to 8 tokens per word for Indic scripts; Sarvam's models aim for a much more efficient 1.4-2.1 tokens per word. - Sarvam's latest models include larger 30B and 105B parameter variants, as well as specialized vision and speech models, trained on trillions of India-specific data tokens to handle complexities like "Hinglish." - As part of the government's IndiaAI Mission, Sarvam AI was selected to build the nation's first sovereign large language model, granting it access to a significant compute infrastructure of over 4,000 GPUs to support development. - Beyond language models, the company is developing a full-stack platform that includes "Sarvam Edge" for running models locally on devices without internet, addressing connectivity and privacy concerns. - The startup has also developed Sarvam-Translate, a model built using Google's Gemma 3, to accurately translate across all 22 officially recognized Indian languages while preserving context and document structure.
