I did a search for Indian AI LLMs.
Found articles containing with headlines containing phrases like: "10 Indian LLMs"
In such articles, almost every link to an Indian LLM was either a 404, or a landing page with a lot of claims but nothing that could be tried out, nothing open source, no github.
I researched press releases. The loudest claims of breakthrough progress were from an entity called "MahindraTech". When I went to their website, there was a message on there saying it was not accessible from outside India (I was trying to access it from outside India).
This reminds me of something from before the April 2024 Indian general elections. There was a huge hooplah about how Modi's speeches were going to be simultaneously translated into multiple Indian languages in real-time using AI. I was very excited about it. Nothing ever came of it.
Right now, if you look into the AI scene, there are startups that are excited about offering Indian-hosted DeepSeek. There are media articles lauding how DeepSeek (China's DeepSeek!) is being used as a pre-trained model to do fine-tuning on. I even saw a government minister talking about how DeepSeek would be hosted on cheap government owned infrastructure for Indian startups. They seem to have given up on pre-training an LLM.
The only serious thing I came across was a quote by Narayan Murthy (founder of InfoSys) saying that Indian AI entrepreneurs should come clean and admit that there isn't enough data to train AIs for Indian language translation.
The rest was just hype and vapor.
It's really quite pathetic.
(Note: I haven't posted links because HN rules don't allow that in a post)
What does he know about this, exactly? I managed to translate a hundred English classics into Sanskrit using DeepSeek. The translation quality is quite good, close to Claude-level.
Gemini 2 Flash, which I am not yet satisfied with as far as Sanskrit is concerned, gave me a brilliant English translation of the Hanuman Chalisa. Even managed to give me an iambic pentameter version of it when prompted.
The data exists. The only question is whether training models from scratch is really necessary or not.
reply