Sarvam AI Launches Edge Model to Compete with Google
Indian AI firm Sarvam AI has launched a new edge-optimized model designed to compete with offerings from Google and OpenAI. The model prioritizes privacy and low latency for on-device applications. Sarvam is targeting enterprise and consumer IoT use cases, including smart home and property management.
- The company raised a significant $41 million in a Series A funding round led by Lightspeed Venture Partners, with participation from Peak XV Partners (formerly Sequoia India) and Khosla Ventures. - Sarvam AI was founded by Dr. Vivek Raghavan, who has experience building India's digital public infrastructure like Aadhaar, and Dr. Pratyush Kumar, a key figure in developing open-source AI for Indian languages through AI4Bharat. - The on-device speech recognition model for 10 Indian languages has a memory footprint of about 294MB and processes speech at 8.5x real-time on a Qualcomm Snapdragon 8 Gen 3 chip. - Its translation model is 150 million parameters, takes up 334MB of device space, and supports 110 language pairs, outperforming Meta's NLLB-600M model on the FloRes benchmark across tested Indian languages. - A key feature is custom voice cloning; a new voice can be added to the text-to-speech model using approximately one hour of audio data, fitting within the same 60MB model file. - The company's prior work includes the OpenHathi series, an open-source Hindi large language model built on Meta's Llama2-7B architecture, which performs on par with GPT-3.5 for Indic languages. - The Indian government selected Sarvam AI to build the country's first sovereign foundational model under the IndiaAI Mission, granting the startup access to 4,000 GPUs for development.
