A large language model (LLM) is an AI model trained on huge amounts of text, so that it can read text and write new text: answers, summaries, translations, letters and code. It is the kind of model behind chatbots such as ChatGPT. Large means that it has billions of internal numbers, called parameters, which are adjusted during training. In India, companies are now training such models on Indian languages under the government's national AI programme.
How does it work? Text is first cut into small pieces called tokens, which may be whole words or parts of words. The model is trained on one task, repeated billions of times: given the tokens so far, predict the next one. To do this well across books, websites, news and code, it has to learn grammar, facts and some reasoning. This first stage is called pre-training, and it needs thousands of GPUs for weeks. In a second stage, the model is trained on examples of good answers written by people, and on people's judgements of which of two answers is better. This teaches it to follow instructions. Data labelling companies supply these examples.
When you type a question, the model writes the answer one token at a time. Each new token is chosen from the model's estimate of what is most likely to come next. So the model can write fluent text that is wrong. That is why companies test it with LLM evaluation, and give it their own documents to read with each question.
Why train a model on Indian languages? Most widely used models learned mainly from English text, and they often handle Hindi, Tamil or Odia less well. Many of them also cut text in Indian scripts into more tokens per word. Providers charge by the token, so the same answer can cost more in Hindi than in English. A model trained on Indian languages can be more accurate and cheaper for Indian users. It can also be kept under Indian control, the idea called sovereign AI.
Imagine a district collectorate in Madurai that receives 2,000 grievance petitions a month, most of them in Tamil. Officials use an assistant built on a language model. For each petition, the model writes a two-line summary in English, names the department responsible and drafts a reply in Tamil. An official reads each draft, corrects it and signs it.
Companies rarely use a model directly. They reach it through an LLM gateway, run it on a model deployment platform, and build products on it, such as a conversational AI platform. A foundation model is the general term for any very large model that others build on. An LLM is the kind that works with text.