Available in Hindi, English, Bengali, and Gujarati, the foundational model has been trained on local data and has been designed with the Indian context in mind
Featuring a transformer decoder-only model, Pragna-1B claims to have been trained on 1.25 Bn parameters with a context length of 2048 tokens
Founded by Abhishek Upperwal, Soket AI Labs claims to have built the country’s first open-source multilingual model on top of TinyLlama, which leverages Meta’s Llama 2 architecture
Homegrown GenAI startup Soket AI Labs has launched India’s first open-source multilingual foundational model Pragna-1B in partnership with Google Cloud.
As per CNBC-TV18, the model will be available in four Indian languages including Hindi, English, Bengali, and Gujarati. The foundational model has been trained on local data and has been designed with the Indian context in mind.
Featuring a transformer decoder-only model, Pragna-1B claims to have been trained on 1.25 Bn parameters with a context length of 2048 tokens. The company claims that its model has achieved “effectiveness” comparable to that of similar category models in language processing tasks.
“… By leveraging Google Cloud’s AI Infrastructure, we achieved both efficiency and cost-effectiveness in the development of Pragna-1B… Tailored specifically for vernacular languages, Pragna-1B offers balanced language representation and enables faster and more efficient tokenization suited for organisations seeking optimised operations and enhanced functionality,” said Soket AI Labs’ founder Abhishek Upperwal.
Google Cloud India’s country managing director and vice-president Bikram Singh Bedi added, “We are thrilled to partner with Soket AI Labs to democratise AI innovation in India. Built on Google Cloud, the launch of Pragna-1B marks a pioneering leap in Indian language technology, offering enhanced scalability and efficiency for organisations”.
As per the report, Soket soon also plans to list its AI developer platform on the Google Cloud Marketplace to offer a streamlined experience for fine-tuning models. It plans to also list its Pragna series of models on the Google Vertex AI model registry.
Founded in 2019 by Upperwal, Soket AI Labs claims to have built the country’s first open-source multilingual model on top of TinyLlama, which leverages Meta’s Llama 2 architecture. Part of NVIDIA’s Inception Programme and AWS Activate, the startup claims to be building full-stack AI solutions for enterprises.
With this, Soket Labs has become the latest Indian startup to emerge from the shadows. Last week, another homegrown GenAI platform ‘Hanooman’ went live in more than 98 languages, including 12 Indian languages.
Prior to that, Bhavish Aggarwal-led AI unicorn Krutrim AI also recently announced the release of an Android app for its AI chatbot. Meanwhile, Meta continues to roll out AI chatbots across its various social media platforms including WhatsApp and Facebook.
At the heart of all this is the flourishing Indian AI space which, as per Inc42, is projected to surpass the $17 Bn mark by 2030.
