The world is after attaining AI supremacy. Already a low-cost DeepSeek –R1 has caused a global sensation. India too proposes to develop indigenous AI models. India intends to support compute power by a stockpile of 18,693 GPUs. There is a proposal of 40 per cent subsidy to developers. It will reduce per hour computing costs.

The state-of-the art global models too suffer from latency and slow response time. They are less efficient than SLMs. LLMs are good performers. Distilled models (DMs) stand in-between. They are relatively less efficient than SLMs.

India cannot remain confined to one type of model. India needs foundation models and LLMs for advanced research in defence, national security and atmospheric studies- – to predict adverse national phenomena, DMS may be used. To answer a query of a farmer in his native language, a model of small and medium size having NLP capabilities could be used.

Generative AI models are capital intensive. Using GPT-4 for finding chemists selling surgical masks is a waste of resources.

LLMs such as Llama-2 and DeepSeek follow ‘open weight'(OW) system of disclosure. It permits users to fine-tune the parent model for customized requirements. OW also enables researchers to test fairness and safety features of a model.

OS models such as Mistral and Falcon not only disclose weights and codes but also information on datasets. Users can do unlimited modification to the parent model and create new models. India’s A14 Bharat model is a pure OS model.

Open Weight models do not disclose training data. They can face lawsuits in markets.