Made-in-India LLMs and where Dhenu fits

Where does Dhenu fit among the made-in-India LLMs? I have been asked that question often enough over the past month, over coffee and on panels, that it deserves a written answer.

I want to be specific and not polemic. The space has real work happening in it, and it has a lot of marketing happening over the real work. Separating the two is useful for anyone choosing what to build or what to fund. The work falls into four categories.

Category 1: indigenous foundation models trained from scratch

The teams doing this seriously: AI4Bharat (IIT Madras) with IndicTrans and the broader Indic foundation-model work, Sarvam AI with their multilingual models, Ola’s Krutrim, Tech Mahindra’s Project Indus, and parts of what TCS Research and Reliance are doing.

These are real efforts with real model checkpoints. They are not finished. Training a competitive foundation model at the scale that matters (multibillion parameters, trillion-token corpora, the engineering required for stable training at that scale) is hard, expensive, and requires sustained commitment that most attempts will not finish.

The fair reading: a year ago there were a lot of pitch decks and few weights. Today there are working models, none yet at GPT-4-class capability on Indic languages, several at the level where they can be useful primitives for downstream applications. The trajectory is up. The endpoint, if the funding holds, is meaningful indigenous capability over the next 2-3 years.

None of this is a shortcut, though. The teams that have published the most useful artefacts (AI4Bharat in particular) have been at it for years, and the teams releasing newcomers every quarter without sustained underlying corpus and training infrastructure are competing for a different prize.

Category 2: fine-tunes of open foundation models on Indic data

This is where most of the visible activity is. Take Llama, Qwen, Mistral, Gemma, fine-tune on Indian data, release on Hugging Face. Some of these are genuinely useful. Some are not.

The useful ones share a property: the team behind them invested in the data as seriously as the fine-tune pass. The fine-tune itself is a few hundred GPU-hours, mechanically straightforward, and available to anyone with the right tooling. The data that makes the result worth using takes months or years of curation. That is where the gap is.

Dhenu’s earlier checkpoints sit in this category. Dhenu 1.0 was a fine-tune over Sarvam’s OpenHathi (an Indic-adapted Llama). Dhenu Llama 3 is a fine-tune over Llama 3 8B. The work that made these useful was the years spent curating Indian agricultural content, building the agronomy-aware retrieval layer, and tuning the multilingual voice pipeline around the model. The fine-tune itself was the easy bit.

The evaluation of this category is simple: a fine-tune is a starting point, not a product. A vendor presenting a fine-tune as a finished product is selling the easy 10% and asking the customer to discover the missing 90% in production.

Category 3: domain models with structured agronomy / legal / medical / financial grounding

This is the category Dhenu eventually graduated into and where I think the long-term action is.

A general Indic LLM, however good, will not handle agricultural compliance correctly. It has no way of knowing that a specific pesticide chemistry is banned in one Indian state and legal in the next, or that the extension publication it learned from in 2019 was superseded by a 2023 amendment. The names farmers use for pests, which vary between districts of a single state, are mostly absent from its training data altogether.

Domain models are the ones with both the language ability and the structured knowledge backbone. Dhenu’s structured agronomy graph is what makes Dhenu different from a generic Indic fine-tune. The same pattern is playing out in legal AI (where Indian-jurisdiction-aware models are starting to matter), medical AI (where regional drug formulary differences matter), and financial AI (where SEBI compliance is its own retrieval problem).

This category is where vertical-specific Indian AI is going to deliver. The general foundation model is necessary infrastructure. The domain layer on top is where the business value sits.

Category 4: branding without substance

This category is the one we should be polite about and unsparing about simultaneously. There are products being marketed as “made-in-India AI” or “Bharat LLM” that are, on inspection, light wrappers around OpenAI or Anthropic API calls with a prompt that says “respond in Hindi.” Nothing about the model is indigenous, and nothing in the system was trained for India. The output quality is the same as any other API consumer’s, with the added cost of an unnecessary middle layer.

This category will collapse on contact with the production realities described above (compliance, dialect, regional taxonomy, last-mile voice). It is not a sustainable shape. The teams in this category will either pivot into doing the actual work or fold.

The reason to call this out is that it muddies the conversation around the real work happening in categories 1 through 3. Buyers cannot tell the categories apart easily. Funders are sometimes happy not to. The result is that the indigenous-AI conversation gets cheaper and noisier than it should be.

Where Dhenu fits

Dhenu is a category-3 product. We are a domain-specific model with a structured agronomy backbone. We sit on top of open foundation models (currently the Llama family with our own fine-tunes), not on top of API calls. We are publishing weights and methodology so that the field can compose with what we are doing, not so we can sell access to a black box.

We are not in category 1 because we are not training foundation models from scratch. We rely on the work in category 1 (AI4Bharat’s Indic infrastructure, Sarvam’s OpenHathi, Meta’s Llama base) and we believe that division of labour is healthy. The teams doing foundation work and the teams doing domain work should be coordinating, not duplicating.

What we want from the made-in-India LLM conversation is more rigor in distinguishing the categories, and more patience for the real work happening in category 1. The shape of the next two years in Indian AI depends on whether the buyers, funders, and policymakers in this space can make these distinctions clearly. The current moment is not encouraging on that front. The teams doing serious work, in all three serious categories, are getting noised over by louder ones.

Some of this clears itself up as production deployments accumulate. Until then, the categories above are the lens I would suggest.