Soket AI to Build India’s First 7B Open-Source Indic LLM
Soket AI Labs, led by CEO Abhishek Upperwal, is embarking on a monumental project to create India’s first 7 billion parameter open-source Indic language model (LLM), with plans to scale it to 120 billion parameters. This ambitious endeavor, part of Project EKA, will make use of region-specific datasets and open-source methodologies to serve key sectors such as defence, healthcare, and education.
Phased Approach to Build a Sovereign AI Model
Soket’s approach to building the LLM is incremental, starting with smaller models (1-2 billion parameters) to test architecture and data alignment. By the 10th month, the goal is to scale up the model to 120 billion parameters. This ambitious vision aims to address India's needs for culturally accurate, secure, and high-performing AI models, especially in sectors like defence where foreign models, including those from China, pose geopolitical risks.
Also Read: Nvidia & Perplexity Boost AI Models in Europe, Middle East
Innovative Data Strategy and Tech Partnerships
Soket’s data strategy focuses heavily on India's underrepresented Indic languages. The team is leveraging advanced techniques like Optical Character Recognition (OCR), Automatic Speech Recognition (ASR), and synthetic data generation through translation. Partnering with institutions like IIT Gandhinagar, Soket is working to digitize and classify a wealth of Indic content, including government records and educational materials, ensuring that the language data remains authentic and culturally relevant.
AI's Local Impact: Moving Beyond Traditional Models
Unlike global AI models that often ignore nuances in languages like Hindi, Soket aims to correct these flaws, providing AI tools that not only understand but respect local dialects and cultural context. With substantial government support, Soket is set to provide India with its own sovereign AI system, reducing reliance on foreign models and accelerating India’s path to AI independence.
