> talent pool

Capital though ;)

[I am a grad student here in reinforcement learning]

Anyways, among all the VC/made-at-home driven snake oil, I'd say you should look at sarvam.ai, they are the most focussed and no-nonsense group. They have a few good from-scratch models (I believe upto 7B or 14B), as well as a few llama finetunes. Their API is pretty good.

The main thing folks here are attempting are to get LLMs good at local indian languages (and I don't mean hindi). I don't think people see a value in creating an "indigenous llama" that doesn't have that property. For this, the main bottleneck is data (relatively speaking, there is zero data in those languages on the internet), so there's a team AI4Bharat whose main job is curating datasets good enough to get stuff like _translation_ and other NLP benchmarks working well. LLMs too, for which they work with sarvam frequently.