On "Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline"

The real bet here isn't whether Indic language models work—it's whether open pretraining at scale can actually compete with the closed-model stranglehold on non-English languages. Two point eight trillion tokens is genuinely substantial, and an open pipeline matters more than the model itself: it means someone else can iterate, someone can adapt it for their market, someone doesn't have to negotiate with OpenAI or Anthropic to serve billions of speakers. For practitioners building in these languages, this flips the power dynamic from "we have to use what's available" to "we have options, and we can build on this." The catch is always adoption and maintenance—open weights mean nothing if the community isn't actually using it and contributing back, which is the graveyard where most language-specific LLMs go to die.