India’s Sarvam launches new large language models to compete with US and Chinese AI systems in open-source market
New Delhi: Indian AI lab Sarvam on Tuesday showed their new language models at a big meeting in India’s capital. The company wants open-source AI that costs less than what bigger companies from America and China offer.
Sarvam says they made new models with 30-billion and 105-billion parts called parameters. These are much bigger than their old 2-billion model from last year. They also made a model that turns text into speech and another that turns speech into text.
The 30-billion model can chat in real-time and understand 32,000 words at once. The bigger 105-billion model can do harder tasks with 128,000 words. These models use less computing power because they only use part of their total parts at any time.
Sarvam trained these new models from the ground up. They used 16 trillion words of text for the smaller model and even more for the bigger one. The training included many Indian languages that people speak in India.
The company plans to share their 30-billion and 105-billion models with everyone for free. They want to make special tools for writing computer code and help businesses use AI. Sarvam started in 2023 and got more than $50 million from investors who believe in their work.
Sarvam says they made new models with 30-billion and 105-billion parts called parameters. These are much bigger than their old 2-billion model from last year. They also made a model that turns text into speech and another that turns speech into text.
The 30-billion model can chat in real-time and understand 32,000 words at once. The bigger 105-billion model can do harder tasks with 128,000 words. These models use less computing power because they only use part of their total parts at any time.
Sarvam trained these new models from the ground up. They used 16 trillion words of text for the smaller model and even more for the bigger one. The training included many Indian languages that people speak in India.
The company plans to share their 30-billion and 105-billion models with everyone for free. They want to make special tools for writing computer code and help businesses use AI. Sarvam started in 2023 and got more than $50 million from investors who believe in their work.